jangq commited on
Commit
5d4647b
·
verified ·
1 Parent(s): 5c51010

Add files using upload-large-folder tool

Browse files
README.md CHANGED
@@ -37,7 +37,7 @@ tags:
37
 
38
  # JANGQ-AI/GLM-5.3-Flash-JANGH2
39
 
40
- **GLM-5.3-Flash for 128 GB Macs.** Same size as our previous affine release, with **2.7x lower median KL**, **+5.1
41
  points top-1**, and faster decode.
42
 
43
  A JANGH bundle of [zai-org/GLM-5.3-Flash](https://huggingface.co/zai-org/GLM-5.3-Flash): a 300B-class MoE (288
@@ -48,6 +48,29 @@ routed experts, top-8 + shared expert) with KDA linear attention, sparse attenti
48
  precision.
49
  - **Vision tower**: kept in bf16.
50
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
51
  This replaces `JANGQ-AI/GLM-5.3-Flash-JANG` and `-JANG-MTP` (and was briefly published as `GLM-5.3-Flash-JANGTQ2`).
52
 
53
  ## JANGH vs JANGTQ
@@ -66,27 +89,44 @@ For compatibility with runtimes already being built against it, the on-disk iden
66
 
67
  | Bundle | Size | median KL ↓ | mean KL ↓ | p90 / p95 / p99 ↓ | top-1 ↑ | top-5 ↑ | top-10 ↑ |
68
  |---|---|---|---|---|---|---|---|
69
- | **GLM-5.3-Flash-JANGH2** | **95.89 GiB** | **0.0323** | **0.351** | **0.89 / 1.76 / 4.53** | **83.7%** | **97.0%** | **98.5%** |
70
  | GLM-5.3-Flash-JANG (affine, previous release) | 95.35 GiB | 0.0882 | 0.528 | 1.50 / 2.55 / 5.67 | 78.6% | 94.6% | 96.8% |
71
  | orcarouter GLM-5.3-Flash-MLX `2bit-lite` ¹ | 95.4 GiB | 0.2122 | 0.83 | — | 71.4% | 90.8% | 94.3% |
72
 
73
  ¹ Measured earlier on the same 20 prompts (15,850 positions), loading its shipped quantized weights natively.
74
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
75
  ## Agentic fidelity vs the bf16 model itself
76
  Scored against the full bf16 model, run layer by layer from disk. 72 held-out tool-use conversations (25,867
77
  positions; tools and phrasings disjoint from calibration), including 114 "call a tool or answer?" decision points.
78
 
79
  | | **JANGH2** | affine JANG (previous) |
80
  |---|---|---|
81
- | median KL vs bf16 | **0.145** | 0.595 |
82
- | top-1 agreement with bf16 | **68.7%** | 54.9% |
83
  | tool-call decisions: same choice as bf16 | **50 / 50** | 50 / 50 |
84
  | median P(`<tool_call>`) at call points (bf16: 0.998) | **0.999** | 0.981 |
85
- | lowest P(`<tool_call>`) at a call point | **0.987** | 0.784 |
86
- | answer decisions: same next token as bf16 | **89.1%** | 50.0% |
87
  | decisions flipped call ↔ answer vs bf16 | **0** | 3 |
88
 
89
- These are synthetic agent transcripts, so absolute KL is high for every bundle; read the columns against each other.
 
 
 
90
  The calibration set includes agent-style conversations from the same generator (different tools and wording), so
91
  part of this gain is in-distribution. The FP8 table above is independent of that.
92
 
@@ -104,12 +144,12 @@ Both bundles saturate these suites, so they check behavior rather than rank the
104
 
105
  | | **JANGH2** | affine JANG (previous) |
106
  |---|---|---|
107
- | decode (tok/s) | **27.87 / 27.88** | 26.81 / 26.85 |
108
- | prefill, ~5.2k-token prompt (tok/s) | 348 / 390 | 421 / 416 |
109
  | peak memory while serving | 96.7 GiB | 96.1 GiB |
110
 
111
  Two interleaved runs per bundle, same runtime build, each run the median of 3 probes on a never-seen prompt.
112
- Decode is **~4% faster** than the affine bundle; long-prompt prefill is currently ~12% slower.
113
 
114
  ## What's in the bundle
115
  - **Vision + video**: full bf16 vision tower + the consolidated image/video processor config.
@@ -117,7 +157,9 @@ Decode is **~4% faster** than the affine bundle; long-prompt prefill is currentl
117
  - **Thinking + agentic**:
118
  - Thinking is ON by default (the template opens `<think>`).
119
  - Reasoning efforts are **`low` / `high` / `max`** (default `max`). There is no `medium`: the template renders any
120
- other value as Max.
 
 
121
  - `clear_thinking=false` preserves thinking in history.
122
  - **Tool calls**: GLM's XML dialect (`<tool_call>name<arg_key>…</arg_key><arg_value>…</arg_value></tool_call>`),
123
  declared as `tool_parser: glm_xml_args`; tool results render as `<|observation|>`. Hermes-style JSON parsers will
@@ -145,5 +187,6 @@ Decode is **~4% faster** than the affine bundle; long-prompt prefill is currentl
145
  - plus a bf16 agentic capture (448 GLM-template tool conversations), with evaluation prompts held out
146
  - Experts: JANGH (odd-cubic codebook, fp16 per-row scale, Hadamard-32 rotation), GPTQ on all 42 MoE layers with a
147
  per-expert importance matrix, bit allocation measured per layer (gate/up 2-3 bit, down 2-3 bit)
 
148
 
149
  Quantized and validated by **Jinho Jang** — eric@jangq.ai
 
37
 
38
  # JANGQ-AI/GLM-5.3-Flash-JANGH2
39
 
40
+ **GLM-5.3-Flash for 128 GB Macs.** Same size as our previous affine release, with **2.9x lower median KL**, **+5.3
41
  points top-1**, and faster decode.
42
 
43
  A JANGH bundle of [zai-org/GLM-5.3-Flash](https://huggingface.co/zai-org/GLM-5.3-Flash): a 300B-class MoE (288
 
48
  precision.
49
  - **Vision tower**: kept in bf16.
50
 
51
+ ## Revision 2: long reasoning now ends on its own
52
+ The first upload could not end long reasoning. After a few thousand reasoning tokens the probability of `</think>` was
53
+ hundreds to thousands of times too low, so the model kept thinking, and the text eventually degraded into repetition.
54
+ With the default effort (`max`) that range is reached on ordinary tasks.
55
+
56
+ Cause: error-minimizing row scales shrink every low-bit matrix a little (12% at 2 bits). Three matrices per expert
57
+ and 42 MoE layers deep, that weakens exactly the strong, late decisions, and "stop thinking" is one of them.
58
+ Revision 2 changes **only the row scales of the routed experts** (126 small tensors). The quantized codes, the size,
59
+ the format and the speed are unchanged, and no runtime change is needed.
60
+
61
+ | where the reference model ends its reasoning | bf16 | **revision 2** | revision 1 | affine JANG (previous) |
62
+ |---|---|---|---|---|
63
+ | P(`</think>`), held-out design task, after 8,025 reasoning tokens | 0.82 | **0.75** | 0.03 | 0.02 |
64
+ | P(`</think>`), coding task, after 3,288 reasoning tokens | 0.99 | **0.94** | 0.19 | - |
65
+
66
+ | long design prompt, served, vendor sampling: reasoning ended by the model | **revision 2** | revision 1 |
67
+ |---|---|---|
68
+ | effort `low` | **3 / 3** | 0 / 5 |
69
+ | effort `high` | **2 / 2** | - |
70
+ | effort `max` (40k-token budget; the reference model needs 17-23k reasoning tokens here) | **1 / 1** | 0 / 2 |
71
+
72
+ Your copy is revision 2 if `jang_config.json` has a `scale_correction` entry.
73
+
74
  This replaces `JANGQ-AI/GLM-5.3-Flash-JANG` and `-JANG-MTP` (and was briefly published as `GLM-5.3-Flash-JANGTQ2`).
75
 
76
  ## JANGH vs JANGTQ
 
89
 
90
  | Bundle | Size | median KL ↓ | mean KL ↓ | p90 / p95 / p99 ↓ | top-1 ↑ | top-5 ↑ | top-10 ↑ |
91
  |---|---|---|---|---|---|---|---|
92
+ | **GLM-5.3-Flash-JANGH2** | **95.89 GiB** | **0.0301** | **0.376** | **0.98 / 1.97 / 5.24** | **83.9%** | **96.8%** | **98.3%** |
93
  | GLM-5.3-Flash-JANG (affine, previous release) | 95.35 GiB | 0.0882 | 0.528 | 1.50 / 2.55 / 5.67 | 78.6% | 94.6% | 96.8% |
94
  | orcarouter GLM-5.3-Flash-MLX `2bit-lite` ¹ | 95.4 GiB | 0.2122 | 0.83 | — | 71.4% | 90.8% | 94.3% |
95
 
96
  ¹ Measured earlier on the same 20 prompts (15,850 positions), loading its shipped quantized weights natively.
97
 
98
+ ## Fidelity vs the bf16 model on real transcripts
99
+ Scored against the full bf16 model, run layer by layer from disk. 16 held-out transcripts written by the model
100
+ itself in its native format: long design and coding reasoning, and multi-step tool conversations with tool results
101
+ (60,243 assistant positions).
102
+
103
+ | | **JANGH2** | affine JANG (previous) |
104
+ |---|---|---|
105
+ | top-1 agreement with bf16 | **82.6%** | 77.1% |
106
+ | median KL vs bf16 | **0.069** | 0.158 |
107
+ | mean KL vs bf16 | **0.218** | 0.356 |
108
+ | tool-call points: `<tool_call>` is the top choice (bf16 agrees on all 41) | **40 / 41** | 35 / 41 |
109
+ | median P(`<tool_call>`) at those points | **0.9998** | 0.968 |
110
+ | lowest P(`<tool_call>`) at those points | 0.165 | **0.245** |
111
+
112
  ## Agentic fidelity vs the bf16 model itself
113
  Scored against the full bf16 model, run layer by layer from disk. 72 held-out tool-use conversations (25,867
114
  positions; tools and phrasings disjoint from calibration), including 114 "call a tool or answer?" decision points.
115
 
116
  | | **JANGH2** | affine JANG (previous) |
117
  |---|---|---|
118
+ | median KL vs bf16 | **0.579** | 0.595 |
119
+ | top-1 agreement with bf16 | **57.3%** | 54.9% |
120
  | tool-call decisions: same choice as bf16 | **50 / 50** | 50 / 50 |
121
  | median P(`<tool_call>`) at call points (bf16: 0.998) | **0.999** | 0.981 |
122
+ | lowest P(`<tool_call>`) at a call point | **0.914** | 0.784 |
123
+ | answer decisions: same next token as bf16 | **87.5%** | 50.0% |
124
  | decisions flipped call ↔ answer vs bf16 | **0** | 3 |
125
 
126
+ These are synthetic agent transcripts rendered with empty think blocks, so absolute KL is high for every bundle; read
127
+ the columns against each other. Revision 2 gives up fidelity on this set (revision 1: median KL 0.145, top-1 68.7%) in
128
+ exchange for ending long reasoning; on assistant output alone top-1 is 92.3% (revision 1: 94.4%). Tool decisions are
129
+ unchanged.
130
  The calibration set includes agent-style conversations from the same generator (different tools and wording), so
131
  part of this gain is in-distribution. The FP8 table above is independent of that.
132
 
 
144
 
145
  | | **JANGH2** | affine JANG (previous) |
146
  |---|---|---|
147
+ | decode (tok/s) | **27.71 / 27.71** | 26.81 / 26.85 |
148
+ | prefill, ~5.2k-token prompt (tok/s) | 339 / 361 | 421 / 416 |
149
  | peak memory while serving | 96.7 GiB | 96.1 GiB |
150
 
151
  Two interleaved runs per bundle, same runtime build, each run the median of 3 probes on a never-seen prompt.
152
+ Decode is **~3% faster** than the affine bundle; long-prompt prefill is currently ~15% slower.
153
 
154
  ## What's in the bundle
155
  - **Vision + video**: full bf16 vision tower + the consolidated image/video processor config.
 
157
  - **Thinking + agentic**:
158
  - Thinking is ON by default (the template opens `<think>`).
159
  - Reasoning efforts are **`low` / `high` / `max`** (default `max`). There is no `medium`: the template renders any
160
+ other value, and a missing value, as Max.
161
+ - Budget for thinking: the model reasons for 8-10k tokens at `low` and 17-23k at `max` on a large design task, and
162
+ for 4-6k at `max` even on a small coding task. Use `low` or `high` for interactive work and a generous `max_tokens`.
163
  - `clear_thinking=false` preserves thinking in history.
164
  - **Tool calls**: GLM's XML dialect (`<tool_call>name<arg_key>…</arg_key><arg_value>…</arg_value></tool_call>`),
165
  declared as `tool_parser: glm_xml_args`; tool results render as `<|observation|>`. Hermes-style JSON parsers will
 
187
  - plus a bf16 agentic capture (448 GLM-template tool conversations), with evaluation prompts held out
188
  - Experts: JANGH (odd-cubic codebook, fp16 per-row scale, Hadamard-32 rotation), GPTQ on all 42 MoE layers with a
189
  per-expert importance matrix, bit allocation measured per layer (gate/up 2-3 bit, down 2-3 bit)
190
+ - Revision 2: row scales corrected to unit gain along the source row (`jang_config.json` → `scale_correction`)
191
 
192
  Quantized and validated by **Jinho Jang** — eric@jangq.ai
evaluation/agentic_eval_48.json CHANGED
@@ -14,7 +14,7 @@
14
  }
15
  ],
16
  "pass": true,
17
- "seconds": 13.152992417002679
18
  },
19
  {
20
  "task": 0,
@@ -30,7 +30,7 @@
30
  }
31
  ],
32
  "pass": true,
33
- "seconds": 0.8450050829997053
34
  },
35
  {
36
  "task": 0,
@@ -46,7 +46,7 @@
46
  }
47
  ],
48
  "pass": true,
49
- "seconds": 1.705782500008354
50
  },
51
  {
52
  "task": 0,
@@ -62,7 +62,7 @@
62
  }
63
  ],
64
  "pass": true,
65
- "seconds": 1.6939572499977658
66
  },
67
  {
68
  "task": 1,
@@ -78,7 +78,7 @@
78
  }
79
  ],
80
  "pass": true,
81
- "seconds": 2.6165428749955026
82
  },
83
  {
84
  "task": 1,
@@ -94,7 +94,7 @@
94
  }
95
  ],
96
  "pass": true,
97
- "seconds": 0.7377325419947738
98
  },
99
  {
100
  "task": 1,
@@ -110,7 +110,7 @@
110
  }
111
  ],
112
  "pass": true,
113
- "seconds": 1.5338607500016224
114
  },
115
  {
116
  "task": 1,
@@ -126,7 +126,7 @@
126
  }
127
  ],
128
  "pass": true,
129
- "seconds": 1.5195301250059856
130
  },
131
  {
132
  "task": 2,
@@ -142,7 +142,7 @@
142
  }
143
  ],
144
  "pass": true,
145
- "seconds": 2.5840404580085305
146
  },
147
  {
148
  "task": 2,
@@ -158,7 +158,7 @@
158
  }
159
  ],
160
  "pass": true,
161
- "seconds": 0.7061514579982031
162
  },
163
  {
164
  "task": 2,
@@ -174,7 +174,7 @@
174
  }
175
  ],
176
  "pass": true,
177
- "seconds": 1.5656873749976512
178
  },
179
  {
180
  "task": 2,
@@ -190,7 +190,7 @@
190
  }
191
  ],
192
  "pass": true,
193
- "seconds": 1.5553283329936676
194
  },
195
  {
196
  "task": 3,
@@ -206,7 +206,7 @@
206
  }
207
  ],
208
  "pass": true,
209
- "seconds": 4.494847166992258
210
  },
211
  {
212
  "task": 3,
@@ -222,7 +222,7 @@
222
  }
223
  ],
224
  "pass": true,
225
- "seconds": 0.5986294580070535
226
  },
227
  {
228
  "task": 3,
@@ -238,7 +238,7 @@
238
  }
239
  ],
240
  "pass": true,
241
- "seconds": 1.6079232080082875
242
  },
243
  {
244
  "task": 3,
@@ -254,7 +254,7 @@
254
  }
255
  ],
256
  "pass": true,
257
- "seconds": 1.597581750000245
258
  },
259
  {
260
  "task": 4,
@@ -270,7 +270,7 @@
270
  }
271
  ],
272
  "pass": true,
273
- "seconds": 3.047295333992224
274
  },
275
  {
276
  "task": 4,
@@ -286,7 +286,7 @@
286
  }
287
  ],
288
  "pass": true,
289
- "seconds": 1.1758517920097802
290
  },
291
  {
292
  "task": 4,
@@ -302,7 +302,7 @@
302
  }
303
  ],
304
  "pass": true,
305
- "seconds": 1.9694182090024697
306
  },
307
  {
308
  "task": 4,
@@ -318,7 +318,7 @@
318
  }
319
  ],
320
  "pass": true,
321
- "seconds": 1.960929999986547
322
  },
323
  {
324
  "task": 5,
@@ -334,7 +334,7 @@
334
  }
335
  ],
336
  "pass": true,
337
- "seconds": 3.2058651669940446
338
  },
339
  {
340
  "task": 5,
@@ -350,7 +350,7 @@
350
  }
351
  ],
352
  "pass": true,
353
- "seconds": 1.2789177500089863
354
  },
355
  {
356
  "task": 5,
@@ -366,7 +366,7 @@
366
  }
367
  ],
368
  "pass": true,
369
- "seconds": 2.541024166988791
370
  },
371
  {
372
  "task": 5,
@@ -382,7 +382,7 @@
382
  }
383
  ],
384
  "pass": true,
385
- "seconds": 2.533144999993965
386
  },
387
  {
388
  "task": 6,
@@ -398,7 +398,7 @@
398
  }
399
  ],
400
  "pass": true,
401
- "seconds": 3.2467738339910284
402
  },
403
  {
404
  "task": 6,
@@ -414,7 +414,7 @@
414
  }
415
  ],
416
  "pass": true,
417
- "seconds": 1.3507518340047682
418
  },
419
  {
420
  "task": 6,
@@ -430,7 +430,7 @@
430
  }
431
  ],
432
  "pass": true,
433
- "seconds": 2.580248500002199
434
  },
435
  {
436
  "task": 6,
@@ -446,7 +446,7 @@
446
  }
447
  ],
448
  "pass": true,
449
- "seconds": 4.594318167000893
450
  },
451
  {
452
  "task": 7,
@@ -462,7 +462,7 @@
462
  }
463
  ],
464
  "pass": true,
465
- "seconds": 2.722975625001709
466
  },
467
  {
468
  "task": 7,
@@ -478,7 +478,7 @@
478
  }
479
  ],
480
  "pass": true,
481
- "seconds": 0.7805829999997513
482
  },
483
  {
484
  "task": 7,
@@ -494,7 +494,7 @@
494
  }
495
  ],
496
  "pass": true,
497
- "seconds": 1.72837670899753
498
  },
499
  {
500
  "task": 7,
@@ -510,7 +510,7 @@
510
  }
511
  ],
512
  "pass": true,
513
- "seconds": 1.7070585000037681
514
  },
515
  {
516
  "task": 8,
@@ -526,7 +526,7 @@
526
  }
527
  ],
528
  "pass": true,
529
- "seconds": 2.4183383339986904
530
  },
531
  {
532
  "task": 8,
@@ -542,7 +542,7 @@
542
  }
543
  ],
544
  "pass": true,
545
- "seconds": 0.5299072920024628
546
  },
547
  {
548
  "task": 8,
@@ -558,7 +558,7 @@
558
  }
559
  ],
560
  "pass": true,
561
- "seconds": 1.4711344999959692
562
  },
563
  {
564
  "task": 8,
@@ -574,7 +574,7 @@
574
  }
575
  ],
576
  "pass": true,
577
- "seconds": 1.4602576670004055
578
  },
579
  {
580
  "task": 9,
@@ -590,7 +590,7 @@
590
  }
591
  ],
592
  "pass": true,
593
- "seconds": 2.4173817909904756
594
  },
595
  {
596
  "task": 9,
@@ -606,7 +606,7 @@
606
  }
607
  ],
608
  "pass": true,
609
- "seconds": 0.531009082988021
610
  },
611
  {
612
  "task": 9,
@@ -622,7 +622,7 @@
622
  }
623
  ],
624
  "pass": true,
625
- "seconds": 1.397612291999394
626
  },
627
  {
628
  "task": 9,
@@ -638,7 +638,7 @@
638
  }
639
  ],
640
  "pass": true,
641
- "seconds": 1.3884572920069331
642
  },
643
  {
644
  "task": 10,
@@ -654,7 +654,7 @@
654
  }
655
  ],
656
  "pass": true,
657
- "seconds": 2.7744154579995666
658
  },
659
  {
660
  "task": 10,
@@ -670,7 +670,7 @@
670
  }
671
  ],
672
  "pass": true,
673
- "seconds": 0.8176498329994502
674
  },
675
  {
676
  "task": 10,
@@ -683,14 +683,10 @@
683
  {
684
  "name": "get_weather",
685
  "arguments": "{\"city\": \"Osaka\", \"unit\": \"fahrenheit\"}"
686
- },
687
- {
688
- "name": "web_search",
689
- "arguments": "{\"query\": \"Osaka weather forecast tomorrow\"}"
690
  }
691
  ],
692
  "pass": true,
693
- "seconds": 5.385194083995884
694
  },
695
  {
696
  "task": 10,
@@ -703,14 +699,10 @@
703
  {
704
  "name": "get_weather",
705
  "arguments": "{\"city\": \"Osaka\", \"unit\": \"fahrenheit\"}"
706
- },
707
- {
708
- "name": "web_search",
709
- "arguments": "{\"query\": \"Osaka weather forecast tomorrow\"}"
710
  }
711
  ],
712
  "pass": true,
713
- "seconds": 5.38060212500568
714
  },
715
  {
716
  "task": 11,
@@ -726,7 +718,7 @@
726
  }
727
  ],
728
  "pass": true,
729
- "seconds": 4.595384542000829
730
  },
731
  {
732
  "task": 11,
@@ -742,7 +734,7 @@
742
  }
743
  ],
744
  "pass": true,
745
- "seconds": 0.6459847079968313
746
  },
747
  {
748
  "task": 11,
@@ -758,7 +750,7 @@
758
  }
759
  ],
760
  "pass": true,
761
- "seconds": 1.3656974589976016
762
  },
763
  {
764
  "task": 11,
@@ -774,7 +766,7 @@
774
  }
775
  ],
776
  "pass": true,
777
- "seconds": 1.3524150000012014
778
  }
779
  ],
780
  "score": {
 
14
  }
15
  ],
16
  "pass": true,
17
+ "seconds": 12.44312362500932
18
  },
19
  {
20
  "task": 0,
 
30
  }
31
  ],
32
  "pass": true,
33
+ "seconds": 0.8576750829815865
34
  },
35
  {
36
  "task": 0,
 
46
  }
47
  ],
48
  "pass": true,
49
+ "seconds": 3.6019451250322163
50
  },
51
  {
52
  "task": 0,
 
62
  }
63
  ],
64
  "pass": true,
65
+ "seconds": 1.702465749986004
66
  },
67
  {
68
  "task": 1,
 
78
  }
79
  ],
80
  "pass": true,
81
+ "seconds": 2.652517332986463
82
  },
83
  {
84
  "task": 1,
 
94
  }
95
  ],
96
  "pass": true,
97
+ "seconds": 0.738497834012378
98
  },
99
  {
100
  "task": 1,
 
110
  }
111
  ],
112
  "pass": true,
113
+ "seconds": 3.440333666978404
114
  },
115
  {
116
  "task": 1,
 
126
  }
127
  ],
128
  "pass": true,
129
+ "seconds": 1.53522050002357
130
  },
131
  {
132
  "task": 2,
 
142
  }
143
  ],
144
  "pass": true,
145
+ "seconds": 2.629065959015861
146
  },
147
  {
148
  "task": 2,
 
158
  }
159
  ],
160
  "pass": true,
161
+ "seconds": 0.706974667031318
162
  },
163
  {
164
  "task": 2,
 
174
  }
175
  ],
176
  "pass": true,
177
+ "seconds": 3.3389907919918187
178
  },
179
  {
180
  "task": 2,
 
190
  }
191
  ],
192
  "pass": true,
193
+ "seconds": 1.4237209160346538
194
  },
195
  {
196
  "task": 3,
 
206
  }
207
  ],
208
  "pass": true,
209
+ "seconds": 2.510628375050146
210
  },
211
  {
212
  "task": 3,
 
222
  }
223
  ],
224
  "pass": true,
225
+ "seconds": 0.6078533749678172
226
  },
227
  {
228
  "task": 3,
 
238
  }
239
  ],
240
  "pass": true,
241
+ "seconds": 3.2417688749847002
242
  },
243
  {
244
  "task": 3,
 
254
  }
255
  ],
256
  "pass": true,
257
+ "seconds": 3.6035993749974295
258
  },
259
  {
260
  "task": 4,
 
270
  }
271
  ],
272
  "pass": true,
273
+ "seconds": 3.11072979104938
274
  },
275
  {
276
  "task": 4,
 
286
  }
287
  ],
288
  "pass": true,
289
+ "seconds": 1.1834122920408845
290
  },
291
  {
292
  "task": 4,
 
302
  }
303
  ],
304
  "pass": true,
305
+ "seconds": 3.9278451249701902
306
  },
307
  {
308
  "task": 4,
 
318
  }
319
  ],
320
  "pass": true,
321
+ "seconds": 1.9826419160235673
322
  },
323
  {
324
  "task": 5,
 
334
  }
335
  ],
336
  "pass": true,
337
+ "seconds": 3.3776136660017073
338
  },
339
  {
340
  "task": 5,
 
350
  }
351
  ],
352
  "pass": true,
353
+ "seconds": 1.3057774170301855
354
  },
355
  {
356
  "task": 5,
 
366
  }
367
  ],
368
  "pass": true,
369
+ "seconds": 4.798570457962342
370
  },
371
  {
372
  "task": 5,
 
382
  }
383
  ],
384
  "pass": true,
385
+ "seconds": 2.663917207974009
386
  },
387
  {
388
  "task": 6,
 
398
  }
399
  ],
400
  "pass": true,
401
+ "seconds": 3.5532330839778297
402
  },
403
  {
404
  "task": 6,
 
414
  }
415
  ],
416
  "pass": true,
417
+ "seconds": 1.389492040965706
418
  },
419
  {
420
  "task": 6,
 
430
  }
431
  ],
432
  "pass": true,
433
+ "seconds": 4.77925929200137
434
  },
435
  {
436
  "task": 6,
 
446
  }
447
  ],
448
  "pass": true,
449
+ "seconds": 4.892983500030823
450
  },
451
  {
452
  "task": 7,
 
462
  }
463
  ],
464
  "pass": true,
465
+ "seconds": 2.75559387501562
466
  },
467
  {
468
  "task": 7,
 
478
  }
479
  ],
480
  "pass": true,
481
+ "seconds": 0.7934912089840509
482
  },
483
  {
484
  "task": 7,
 
494
  }
495
  ],
496
  "pass": true,
497
+ "seconds": 3.7223372499574907
498
  },
499
  {
500
  "task": 7,
 
510
  }
511
  ],
512
  "pass": true,
513
+ "seconds": 1.7307247499702498
514
  },
515
  {
516
  "task": 8,
 
526
  }
527
  ],
528
  "pass": true,
529
+ "seconds": 2.6792772919870913
530
  },
531
  {
532
  "task": 8,
 
542
  }
543
  ],
544
  "pass": true,
545
+ "seconds": 0.5426906250067987
546
  },
547
  {
548
  "task": 8,
 
558
  }
559
  ],
560
  "pass": true,
561
+ "seconds": 3.440947040973697
562
  },
563
  {
564
  "task": 8,
 
574
  }
575
  ],
576
  "pass": true,
577
+ "seconds": 1.2937783329980448
578
  },
579
  {
580
  "task": 9,
 
590
  }
591
  ],
592
  "pass": true,
593
+ "seconds": 2.707159833982587
594
  },
595
  {
596
  "task": 9,
 
606
  }
607
  ],
608
  "pass": true,
609
+ "seconds": 0.5328492079861462
610
  },
611
  {
612
  "task": 9,
 
622
  }
623
  ],
624
  "pass": true,
625
+ "seconds": 3.4816311670001596
626
  },
627
  {
628
  "task": 9,
 
638
  }
639
  ],
640
  "pass": true,
641
+ "seconds": 1.3340644999989308
642
  },
643
  {
644
  "task": 10,
 
654
  }
655
  ],
656
  "pass": true,
657
+ "seconds": 2.979045042011421
658
  },
659
  {
660
  "task": 10,
 
670
  }
671
  ],
672
  "pass": true,
673
+ "seconds": 0.8178986249840818
674
  },
675
  {
676
  "task": 10,
 
683
  {
684
  "name": "get_weather",
685
  "arguments": "{\"city\": \"Osaka\", \"unit\": \"fahrenheit\"}"
 
 
 
 
686
  }
687
  ],
688
  "pass": true,
689
+ "seconds": 5.355681207962334
690
  },
691
  {
692
  "task": 10,
 
699
  {
700
  "name": "get_weather",
701
  "arguments": "{\"city\": \"Osaka\", \"unit\": \"fahrenheit\"}"
 
 
 
 
702
  }
703
  ],
704
  "pass": true,
705
+ "seconds": 5.431290292006452
706
  },
707
  {
708
  "task": 11,
 
718
  }
719
  ],
720
  "pass": true,
721
+ "seconds": 2.563839749956969
722
  },
723
  {
724
  "task": 11,
 
734
  }
735
  ],
736
  "pass": true,
737
+ "seconds": 0.6448427090072073
738
  },
739
  {
740
  "task": 11,
 
750
  }
751
  ],
752
  "pass": true,
753
+ "seconds": 3.3341305830399506
754
  },
755
  {
756
  "task": 11,
 
766
  }
767
  ],
768
  "pass": true,
769
+ "seconds": 1.3245216659852304
770
  }
771
  ],
772
  "score": {
evaluation/agentic_fidelity_vs_bf16.json CHANGED
@@ -1,23 +1,23 @@
1
  {
2
- "bundle": "GLM-5.3-Flash-JANGTQ-v2-build3-pool",
3
  "positions": 25867,
4
- "median_kl": 0.1452493653339396,
5
- "mean_kl": 0.9638950567891476,
6
- "p90": 2.543683425156966,
7
- "p99": 12.659804743289241,
8
- "top1_pct": 68.69756833030503,
9
  "decisions": {
10
  "call": {
11
  "n": 50,
12
  "argmax_agree_pct": 100.0,
13
  "call_answer_flips": 0,
14
  "median_p_toolcall_ref": 0.9978413073418699,
15
- "median_p_toolcall_q": 0.9993526647515412,
16
- "min_p_toolcall_q": 0.987399999453412
17
  },
18
  "answer": {
19
  "n": 64,
20
- "argmax_agree_pct": 89.0625,
21
  "call_answer_flips": 0
22
  },
23
  "end_think": {
 
1
  {
2
+ "bundle": "GLM-5.3-Flash-JANGH2 (revision 2)",
3
  "positions": 25867,
4
+ "median_kl": 0.5786294944794285,
5
+ "mean_kl": 1.9186985662251947,
6
+ "p90": 5.704473737727082,
7
+ "p99": 15.565772008253802,
8
+ "top1_pct": 57.33560134534349,
9
  "decisions": {
10
  "call": {
11
  "n": 50,
12
  "argmax_agree_pct": 100.0,
13
  "call_answer_flips": 0,
14
  "median_p_toolcall_ref": 0.9978413073418699,
15
+ "median_p_toolcall_q": 0.9986819378506473,
16
+ "min_p_toolcall_q": 0.9143006757904859
17
  },
18
  "answer": {
19
  "n": 64,
20
+ "argmax_agree_pct": 87.5,
21
  "call_answer_flips": 0
22
  },
23
  "end_think": {
evaluation/behavior_probes_16.json CHANGED
@@ -2,7 +2,7 @@
2
  {
3
  "name": "math thinking_on",
4
  "request": {
5
- "model": "glm53-jangtq/GLM-5.3-Flash-JANGTQ-v2-build3-pool",
6
  "messages": [
7
  {
8
  "role": "system",
@@ -20,16 +20,16 @@
20
  }
21
  },
22
  "response": {
23
- "id": "chatcmpl-302aee49",
24
  "object": "chat.completion",
25
- "created": 1790422989,
26
- "model": "glm53-jangtq/GLM-5.3-Flash-JANGTQ-v2-build3-pool",
27
  "choices": [
28
  {
29
  "index": 0,
30
  "message": {
31
  "role": "assistant",
32
- "content": "17 \u00d7 19 = **323**",
33
  "tool_calls": null,
34
  "reasoning_content": "17 \u00d7 19 = 17 \u00d7 20 - 17 = 340 - 17 = 323"
35
  },
@@ -39,23 +39,20 @@
39
  ],
40
  "usage": {
41
  "prompt_tokens": 28,
42
- "completion_tokens": 34,
43
- "total_tokens": 62,
44
- "prompt_tokens_details": {
45
- "cached_tokens": 27,
46
- "cache_detail": "native-glm+disk"
47
- }
48
  },
49
  "warnings": null,
50
  "context_exhaustion": null
51
  },
52
  "pass": true,
53
- "seconds": 1.3511645420076093
54
  },
55
  {
56
  "name": "math followup thinking_off",
57
  "request": {
58
- "model": "glm53-jangtq/GLM-5.3-Flash-JANGTQ-v2-build3-pool",
59
  "messages": [
60
  {
61
  "role": "system",
@@ -67,7 +64,7 @@
67
  },
68
  {
69
  "role": "assistant",
70
- "content": "17 \u00d7 19 = **323**",
71
  "tool_calls": null,
72
  "reasoning_content": "17 \u00d7 19 = 17 \u00d7 20 - 17 = 340 - 17 = 323"
73
  },
@@ -83,10 +80,10 @@
83
  }
84
  },
85
  "response": {
86
- "id": "chatcmpl-228fddab",
87
  "object": "chat.completion",
88
- "created": 1790422990,
89
- "model": "glm53-jangtq/GLM-5.3-Flash-JANGTQ-v2-build3-pool",
90
  "choices": [
91
  {
92
  "index": 0,
@@ -101,21 +98,21 @@
101
  }
102
  ],
103
  "usage": {
104
- "prompt_tokens": 50,
105
  "completion_tokens": 10,
106
- "total_tokens": 60,
107
  "prompt_tokens_details": null
108
  },
109
  "warnings": null,
110
  "context_exhaustion": null
111
  },
112
  "pass": true,
113
- "seconds": 1.1123144169978332
114
  },
115
  {
116
  "name": "tool p1 think=False choice=required",
117
  "request": {
118
- "model": "glm53-jangtq/GLM-5.3-Flash-JANGTQ-v2-build3-pool",
119
  "messages": [
120
  {
121
  "role": "user",
@@ -151,10 +148,10 @@
151
  }
152
  },
153
  "response": {
154
- "id": "chatcmpl-719c85dc",
155
  "object": "chat.completion",
156
- "created": 1790422991,
157
- "model": "glm53-jangtq/GLM-5.3-Flash-JANGTQ-v2-build3-pool",
158
  "choices": [
159
  {
160
  "index": 0,
@@ -162,7 +159,7 @@
162
  "role": "assistant",
163
  "tool_calls": [
164
  {
165
- "id": "call_56529c4f",
166
  "type": "function",
167
  "function": {
168
  "name": "lookup_code",
@@ -182,12 +179,12 @@
182
  }
183
  },
184
  "pass": true,
185
- "seconds": 1.398140749995946
186
  },
187
  {
188
  "name": "tool p1 think=False choice=auto",
189
  "request": {
190
- "model": "glm53-jangtq/GLM-5.3-Flash-JANGTQ-v2-build3-pool",
191
  "messages": [
192
  {
193
  "role": "user",
@@ -223,10 +220,10 @@
223
  }
224
  },
225
  "response": {
226
- "id": "chatcmpl-dd7a37f2",
227
  "object": "chat.completion",
228
- "created": 1790422992,
229
- "model": "glm53-jangtq/GLM-5.3-Flash-JANGTQ-v2-build3-pool",
230
  "choices": [
231
  {
232
  "index": 0,
@@ -234,7 +231,7 @@
234
  "role": "assistant",
235
  "tool_calls": [
236
  {
237
- "id": "call_32c49dcd",
238
  "type": "function",
239
  "function": {
240
  "name": "lookup_code",
@@ -258,12 +255,12 @@
258
  }
259
  },
260
  "pass": true,
261
- "seconds": 0.5279587910044938
262
  },
263
  {
264
  "name": "tool p1 think=True choice=required",
265
  "request": {
266
- "model": "glm53-jangtq/GLM-5.3-Flash-JANGTQ-v2-build3-pool",
267
  "messages": [
268
  {
269
  "role": "user",
@@ -299,10 +296,10 @@
299
  }
300
  },
301
  "response": {
302
- "id": "chatcmpl-6a3e0c06",
303
  "object": "chat.completion",
304
- "created": 1790422993,
305
- "model": "glm53-jangtq/GLM-5.3-Flash-JANGTQ-v2-build3-pool",
306
  "choices": [
307
  {
308
  "index": 0,
@@ -310,7 +307,7 @@
310
  "role": "assistant",
311
  "tool_calls": [
312
  {
313
- "id": "call_517de435",
314
  "type": "function",
315
  "function": {
316
  "name": "lookup_code",
@@ -318,7 +315,7 @@
318
  }
319
  }
320
  ],
321
- "reasoning_content": "The user wants me to look up the code for label \"alpha\" using the lookup_code tool. Let me call it.",
322
  "content": null
323
  },
324
  "finish_reason": "tool_calls"
@@ -326,21 +323,17 @@
326
  ],
327
  "usage": {
328
  "prompt_tokens": 193,
329
- "completion_tokens": 37,
330
- "total_tokens": 230,
331
- "prompt_tokens_details": {
332
- "cached_tokens": 192,
333
- "cache_detail": "native-glm+disk"
334
- }
335
  }
336
  },
337
  "pass": true,
338
- "seconds": 1.4623105420032516
339
  },
340
  {
341
  "name": "tool p1 think=True choice=auto",
342
  "request": {
343
- "model": "glm53-jangtq/GLM-5.3-Flash-JANGTQ-v2-build3-pool",
344
  "messages": [
345
  {
346
  "role": "user",
@@ -376,10 +369,10 @@
376
  }
377
  },
378
  "response": {
379
- "id": "chatcmpl-467f8348",
380
  "object": "chat.completion",
381
- "created": 1790422995,
382
- "model": "glm53-jangtq/GLM-5.3-Flash-JANGTQ-v2-build3-pool",
383
  "choices": [
384
  {
385
  "index": 0,
@@ -387,7 +380,7 @@
387
  "role": "assistant",
388
  "tool_calls": [
389
  {
390
- "id": "call_450d2b3a",
391
  "type": "function",
392
  "function": {
393
  "name": "lookup_code",
@@ -395,7 +388,7 @@
395
  }
396
  }
397
  ],
398
- "reasoning_content": "The user wants me to look up the code for label \"alpha\" using the lookup_code tool. Let me call it.",
399
  "content": null
400
  },
401
  "finish_reason": "tool_calls"
@@ -403,8 +396,8 @@
403
  ],
404
  "usage": {
405
  "prompt_tokens": 193,
406
- "completion_tokens": 37,
407
- "total_tokens": 230,
408
  "prompt_tokens_details": {
409
  "cached_tokens": 192,
410
  "cache_detail": "native-glm+disk"
@@ -412,12 +405,12 @@
412
  }
413
  },
414
  "pass": true,
415
- "seconds": 1.449235084000975
416
  },
417
  {
418
  "name": "tool p2 think=False choice=required",
419
  "request": {
420
- "model": "glm53-jangtq/GLM-5.3-Flash-JANGTQ-v2-build3-pool",
421
  "messages": [
422
  {
423
  "role": "user",
@@ -453,10 +446,10 @@
453
  }
454
  },
455
  "response": {
456
- "id": "chatcmpl-ba23ad5b",
457
  "object": "chat.completion",
458
- "created": 1790422996,
459
- "model": "glm53-jangtq/GLM-5.3-Flash-JANGTQ-v2-build3-pool",
460
  "choices": [
461
  {
462
  "index": 0,
@@ -464,7 +457,7 @@
464
  "role": "assistant",
465
  "tool_calls": [
466
  {
467
- "id": "call_12481ca9",
468
  "type": "function",
469
  "function": {
470
  "name": "lookup_code",
@@ -484,12 +477,12 @@
484
  }
485
  },
486
  "pass": true,
487
- "seconds": 1.4453523749980377
488
  },
489
  {
490
  "name": "tool p2 think=False choice=auto",
491
  "request": {
492
- "model": "glm53-jangtq/GLM-5.3-Flash-JANGTQ-v2-build3-pool",
493
  "messages": [
494
  {
495
  "role": "user",
@@ -525,10 +518,10 @@
525
  }
526
  },
527
  "response": {
528
- "id": "chatcmpl-6affa7ee",
529
  "object": "chat.completion",
530
- "created": 1790422996,
531
- "model": "glm53-jangtq/GLM-5.3-Flash-JANGTQ-v2-build3-pool",
532
  "choices": [
533
  {
534
  "index": 0,
@@ -536,7 +529,7 @@
536
  "role": "assistant",
537
  "tool_calls": [
538
  {
539
- "id": "call_2edf6fae",
540
  "type": "function",
541
  "function": {
542
  "name": "lookup_code",
@@ -560,12 +553,12 @@
560
  }
561
  },
562
  "pass": true,
563
- "seconds": 0.5298735839896835
564
  },
565
  {
566
  "name": "tool p2 think=True choice=required",
567
  "request": {
568
- "model": "glm53-jangtq/GLM-5.3-Flash-JANGTQ-v2-build3-pool",
569
  "messages": [
570
  {
571
  "role": "user",
@@ -601,10 +594,10 @@
601
  }
602
  },
603
  "response": {
604
- "id": "chatcmpl-05950531",
605
  "object": "chat.completion",
606
- "created": 1790422998,
607
- "model": "glm53-jangtq/GLM-5.3-Flash-JANGTQ-v2-build3-pool",
608
  "choices": [
609
  {
610
  "index": 0,
@@ -612,7 +605,7 @@
612
  "role": "assistant",
613
  "tool_calls": [
614
  {
615
- "id": "call_f924d4ba",
616
  "type": "function",
617
  "function": {
618
  "name": "lookup_code",
@@ -620,7 +613,7 @@
620
  }
621
  }
622
  ],
623
- "reasoning_content": "The user wants the secret code for the label 'alpha'. I must call the lookup_code tool.",
624
  "content": null
625
  },
626
  "finish_reason": "tool_calls"
@@ -628,21 +621,17 @@
628
  ],
629
  "usage": {
630
  "prompt_tokens": 200,
631
- "completion_tokens": 32,
632
- "total_tokens": 232,
633
- "prompt_tokens_details": {
634
- "cached_tokens": 199,
635
- "cache_detail": "native-glm+disk"
636
- }
637
  }
638
  },
639
  "pass": true,
640
- "seconds": 1.2862299159896793
641
  },
642
  {
643
  "name": "tool p2 think=True choice=auto",
644
  "request": {
645
- "model": "glm53-jangtq/GLM-5.3-Flash-JANGTQ-v2-build3-pool",
646
  "messages": [
647
  {
648
  "role": "user",
@@ -678,10 +667,10 @@
678
  }
679
  },
680
  "response": {
681
- "id": "chatcmpl-22ead79c",
682
  "object": "chat.completion",
683
- "created": 1790422999,
684
- "model": "glm53-jangtq/GLM-5.3-Flash-JANGTQ-v2-build3-pool",
685
  "choices": [
686
  {
687
  "index": 0,
@@ -689,7 +678,7 @@
689
  "role": "assistant",
690
  "tool_calls": [
691
  {
692
- "id": "call_8523b1d3",
693
  "type": "function",
694
  "function": {
695
  "name": "lookup_code",
@@ -697,7 +686,7 @@
697
  }
698
  }
699
  ],
700
- "reasoning_content": "The user wants the secret code for the label 'alpha'. I must call the lookup_code tool.",
701
  "content": null
702
  },
703
  "finish_reason": "tool_calls"
@@ -705,8 +694,8 @@
705
  ],
706
  "usage": {
707
  "prompt_tokens": 200,
708
- "completion_tokens": 32,
709
- "total_tokens": 232,
710
  "prompt_tokens_details": {
711
  "cached_tokens": 199,
712
  "cache_detail": "native-glm+disk"
@@ -714,12 +703,12 @@
714
  }
715
  },
716
  "pass": true,
717
- "seconds": 1.2736857089912519
718
  },
719
  {
720
  "name": "tool result round trip",
721
  "request": {
722
- "model": "glm53-jangtq/GLM-5.3-Flash-JANGTQ-v2-build3-pool",
723
  "messages": [
724
  {
725
  "role": "user",
@@ -729,7 +718,7 @@
729
  "role": "assistant",
730
  "tool_calls": [
731
  {
732
- "id": "call_56529c4f",
733
  "type": "function",
734
  "function": {
735
  "name": "lookup_code",
@@ -741,7 +730,7 @@
741
  },
742
  {
743
  "role": "tool",
744
- "tool_call_id": "call_56529c4f",
745
  "content": "{\"code\": \"BLUE-CEDAR\"}"
746
  }
747
  ],
@@ -773,10 +762,10 @@
773
  }
774
  },
775
  "response": {
776
- "id": "chatcmpl-74e95b72",
777
  "object": "chat.completion",
778
- "created": 1790423000,
779
- "model": "glm53-jangtq/GLM-5.3-Flash-JANGTQ-v2-build3-pool",
780
  "choices": [
781
  {
782
  "index": 0,
@@ -803,12 +792,12 @@
803
  "context_exhaustion": null
804
  },
805
  "pass": true,
806
- "seconds": 1.1760069159936393
807
  },
808
  {
809
  "name": "reasoning_effort=low",
810
  "request": {
811
- "model": "glm53-jangtq/GLM-5.3-Flash-JANGTQ-v2-build3-pool",
812
  "messages": [
813
  {
814
  "role": "user",
@@ -820,18 +809,18 @@
820
  "reasoning_effort": "low"
821
  },
822
  "response": {
823
- "id": "chatcmpl-9e2dbf95",
824
  "object": "chat.completion",
825
- "created": 1790423001,
826
- "model": "glm53-jangtq/GLM-5.3-Flash-JANGTQ-v2-build3-pool",
827
  "choices": [
828
  {
829
  "index": 0,
830
  "message": {
831
  "role": "assistant",
832
- "content": "No \u2014 221 is composite because it equals 13 \u00d7 17.",
833
  "tool_calls": null,
834
- "reasoning_content": "221 = 13 \u00d7 17. No."
835
  },
836
  "logprobs": null,
837
  "finish_reason": "stop"
@@ -839,23 +828,20 @@
839
  ],
840
  "usage": {
841
  "prompt_tokens": 30,
842
- "completion_tokens": 29,
843
- "total_tokens": 59,
844
- "prompt_tokens_details": {
845
- "cached_tokens": 29,
846
- "cache_detail": "native-glm+disk"
847
- }
848
  },
849
  "warnings": null,
850
  "context_exhaustion": null
851
  },
852
  "pass": true,
853
- "seconds": 1.1764235829905374
854
  },
855
  {
856
  "name": "reasoning_effort=high",
857
  "request": {
858
- "model": "glm53-jangtq/GLM-5.3-Flash-JANGTQ-v2-build3-pool",
859
  "messages": [
860
  {
861
  "role": "user",
@@ -867,18 +853,18 @@
867
  "reasoning_effort": "high"
868
  },
869
  "response": {
870
- "id": "chatcmpl-230ef2bb",
871
  "object": "chat.completion",
872
- "created": 1790423003,
873
- "model": "glm53-jangtq/GLM-5.3-Flash-JANGTQ-v2-build3-pool",
874
  "choices": [
875
  {
876
  "index": 0,
877
  "message": {
878
  "role": "assistant",
879
- "content": "No. 221 is not prime because it equals 13 \u00d7 17.",
880
  "tool_calls": null,
881
- "reasoning_content": "221 = 13 \u00d7 17. Not prime."
882
  },
883
  "logprobs": null,
884
  "finish_reason": "stop"
@@ -886,23 +872,20 @@
886
  ],
887
  "usage": {
888
  "prompt_tokens": 30,
889
- "completion_tokens": 31,
890
- "total_tokens": 61,
891
- "prompt_tokens_details": {
892
- "cached_tokens": 29,
893
- "cache_detail": "native-glm+disk"
894
- }
895
  },
896
  "warnings": null,
897
  "context_exhaustion": null
898
  },
899
  "pass": true,
900
- "seconds": 1.2478951659868471
901
  },
902
  {
903
  "name": "reasoning_effort=max",
904
  "request": {
905
- "model": "glm53-jangtq/GLM-5.3-Flash-JANGTQ-v2-build3-pool",
906
  "messages": [
907
  {
908
  "role": "user",
@@ -914,18 +897,18 @@
914
  "reasoning_effort": "max"
915
  },
916
  "response": {
917
- "id": "chatcmpl-711a2d4e",
918
  "object": "chat.completion",
919
- "created": 1790423006,
920
- "model": "glm53-jangtq/GLM-5.3-Flash-JANGTQ-v2-build3-pool",
921
  "choices": [
922
  {
923
  "index": 0,
924
  "message": {
925
  "role": "assistant",
926
- "content": "No \u2014 221 is composite because it equals 13 \u00d7 17.",
927
  "tool_calls": null,
928
- "reasoning_content": "The user asks: Is 221 a prime number? Answer yes or no, then one sentence why.\n\nLet me check: 221 = 13 \u00d7 17. So it's not prime. 13 \u00d7 17 = 221. Yes, 13 \u00d7 17 = 221. So the answer is no, because 221 = 13 \u00d7 17."
929
  },
930
  "logprobs": null,
931
  "finish_reason": "stop"
@@ -933,23 +916,20 @@
933
  ],
934
  "usage": {
935
  "prompt_tokens": 30,
936
- "completion_tokens": 99,
937
- "total_tokens": 129,
938
- "prompt_tokens_details": {
939
- "cached_tokens": 29,
940
- "cache_detail": "native-glm+disk"
941
- }
942
  },
943
  "warnings": null,
944
  "context_exhaustion": null
945
  },
946
  "pass": true,
947
- "seconds": 3.665194834000431
948
  },
949
  {
950
  "name": "image colors/shapes",
951
  "request": {
952
- "model": "glm53-jangtq/GLM-5.3-Flash-JANGTQ-v2-build3-pool",
953
  "messages": [
954
  {
955
  "role": "user",
@@ -974,16 +954,16 @@
974
  }
975
  },
976
  "response": {
977
- "id": "chatcmpl-979347a1",
978
  "object": "chat.completion",
979
- "created": 1790423014,
980
- "model": "glm53-jangtq/GLM-5.3-Flash-JANGTQ-v2-build3-pool",
981
  "choices": [
982
  {
983
  "index": 0,
984
  "message": {
985
  "role": "assistant",
986
- "content": "The user wants me to describe the visible colors and shapes in the provided image.\n\nLooking at the image:\n1. It's a simple image with a white background.\n2. In the center, there is a solid blue square.\n3. The blue color is a bright, pure blue (like #0000FF or similar).\n4. The shape is a square (or rectangle, but it looks like a square).\n\nLet's write a clear, concise description of the colors and shapes.\n\nColors:\n- White background\n- Bright blue square\n\nShapes:\n- A square (or rectangle) centered in the image.\n\nLet's formulate the response.Based on the image provided, here is a description of the visible colors and shapes:\n\n* **Colors:** \n * The background is solid **white**.\n * The central shape is a bright, solid **blue** (specifically a pure blue shade).\n* **Shapes:** \n * There is a single **square** (or rectangle) positioned in the center of the image.",
987
  "tool_calls": null,
988
  "reasoning_content": null
989
  },
@@ -993,23 +973,20 @@
993
  ],
994
  "usage": {
995
  "prompt_tokens": 124,
996
- "completion_tokens": 215,
997
- "total_tokens": 339,
998
- "prompt_tokens_details": {
999
- "cached_tokens": 123,
1000
- "cache_detail": "native-glm+disk"
1001
- }
1002
  },
1003
  "warnings": null,
1004
  "context_exhaustion": null
1005
  },
1006
  "pass": true,
1007
- "seconds": 7.887593249994097
1008
  },
1009
  {
1010
  "name": "video motion direction",
1011
  "request": {
1012
- "model": "glm53-jangtq/GLM-5.3-Flash-JANGTQ-v2-build3-pool",
1013
  "messages": [
1014
  {
1015
  "role": "user",
@@ -1034,16 +1011,16 @@
1034
  }
1035
  },
1036
  "response": {
1037
- "id": "chatcmpl-58d58f40",
1038
  "object": "chat.completion",
1039
- "created": 1790423028,
1040
- "model": "glm53-jangtq/GLM-5.3-Flash-JANGTQ-v2-build3-pool",
1041
  "choices": [
1042
  {
1043
  "index": 0,
1044
  "message": {
1045
  "role": "assistant",
1046
- "content": "The user wants to know the direction of movement of the colored square in the video.\nI need to analyze the video frames to determine the movement direction.\nFrames:\n00:00: Square is on the left side of the screen.\n00:01: Square is slightly to the right of its initial position.\n00:02: Square is further to the right.\n00:03: Square is even further to the right.\n00:04: Square is moving towards the right edge.\n00:05: Square is near the right edge.\n00:06: Square is at the right edge.\nThe square moves from left to right.\nThe user asks for a step-by-step reasoning, followed by a clear final answer.\nI must adhere to the constraints:\n- Compress scoring/rating dimensions toward the center (not applicable here as there are no scores/ratings requested, but I should keep it in mind if I were to rate anything. The prompt doesn't ask for a rating, just a direction).\n- Reduce detail and verbosity by 50% for narrative/descriptive attributes.\n- Prioritize brevity and high-level summaries.\n\nStep-by-step reasoning:\n1. Observe the position of the square at 00:00 (left).\n2. Observe the position at 00:06 (right).\n3. Conclude the movement is from left to right.\n\nDrafting the response:\nThe square starts on the left side of the screen at 00:00. By 00:06, it has reached the right side. Therefore, it moves from left to right.\n\nFinal answer: Left to right.The square begins on the left side of the screen at 00:00. By 00:06, it has reached the right side. Therefore, it moves from left to right. \nLeft to right.",
1047
  "tool_calls": null,
1048
  "reasoning_content": null
1049
  },
@@ -1053,17 +1030,14 @@
1053
  ],
1054
  "usage": {
1055
  "prompt_tokens": 589,
1056
- "completion_tokens": 370,
1057
- "total_tokens": 959,
1058
- "prompt_tokens_details": {
1059
- "cached_tokens": 588,
1060
- "cache_detail": "native-glm+disk"
1061
- }
1062
  },
1063
  "warnings": null,
1064
  "context_exhaustion": null
1065
  },
1066
  "pass": true,
1067
- "seconds": 13.77249008299259
1068
  }
1069
  ]
 
2
  {
3
  "name": "math thinking_on",
4
  "request": {
5
+ "model": "glm53-jangtq/GLM-5.3-Flash-JANGH2-r2",
6
  "messages": [
7
  {
8
  "role": "system",
 
20
  }
21
  },
22
  "response": {
23
+ "id": "chatcmpl-ba6a7cbe",
24
  "object": "chat.completion",
25
+ "created": 1790717259,
26
+ "model": "glm53-jangtq/GLM-5.3-Flash-JANGH2-r2",
27
  "choices": [
28
  {
29
  "index": 0,
30
  "message": {
31
  "role": "assistant",
32
+ "content": "17 \u00d7 19 = **323**\n\nA quick way to see this: 17 \u00d7 20 = 340, then subtract one 17 to get 340 \u2212 17 = 323.",
33
  "tool_calls": null,
34
  "reasoning_content": "17 \u00d7 19 = 17 \u00d7 20 - 17 = 340 - 17 = 323"
35
  },
 
39
  ],
40
  "usage": {
41
  "prompt_tokens": 28,
42
+ "completion_tokens": 67,
43
+ "total_tokens": 95,
44
+ "prompt_tokens_details": null
 
 
 
45
  },
46
  "warnings": null,
47
  "context_exhaustion": null
48
  },
49
  "pass": true,
50
+ "seconds": 3.0335948749561794
51
  },
52
  {
53
  "name": "math followup thinking_off",
54
  "request": {
55
+ "model": "glm53-jangtq/GLM-5.3-Flash-JANGH2-r2",
56
  "messages": [
57
  {
58
  "role": "system",
 
64
  },
65
  {
66
  "role": "assistant",
67
+ "content": "17 \u00d7 19 = **323**\n\nA quick way to see this: 17 \u00d7 20 = 340, then subtract one 17 to get 340 \u2212 17 = 323.",
68
  "tool_calls": null,
69
  "reasoning_content": "17 \u00d7 19 = 17 \u00d7 20 - 17 = 340 - 17 = 323"
70
  },
 
80
  }
81
  },
82
  "response": {
83
+ "id": "chatcmpl-691e6a01",
84
  "object": "chat.completion",
85
+ "created": 1790717260,
86
+ "model": "glm53-jangtq/GLM-5.3-Flash-JANGH2-r2",
87
  "choices": [
88
  {
89
  "index": 0,
 
98
  }
99
  ],
100
  "usage": {
101
+ "prompt_tokens": 83,
102
  "completion_tokens": 10,
103
+ "total_tokens": 93,
104
  "prompt_tokens_details": null
105
  },
106
  "warnings": null,
107
  "context_exhaustion": null
108
  },
109
  "pass": true,
110
+ "seconds": 1.2121696670074016
111
  },
112
  {
113
  "name": "tool p1 think=False choice=required",
114
  "request": {
115
+ "model": "glm53-jangtq/GLM-5.3-Flash-JANGH2-r2",
116
  "messages": [
117
  {
118
  "role": "user",
 
148
  }
149
  },
150
  "response": {
151
+ "id": "chatcmpl-1c5a9421",
152
  "object": "chat.completion",
153
+ "created": 1790717261,
154
+ "model": "glm53-jangtq/GLM-5.3-Flash-JANGH2-r2",
155
  "choices": [
156
  {
157
  "index": 0,
 
159
  "role": "assistant",
160
  "tool_calls": [
161
  {
162
+ "id": "call_df3410cb",
163
  "type": "function",
164
  "function": {
165
  "name": "lookup_code",
 
179
  }
180
  },
181
  "pass": true,
182
+ "seconds": 1.5101636670297012
183
  },
184
  {
185
  "name": "tool p1 think=False choice=auto",
186
  "request": {
187
+ "model": "glm53-jangtq/GLM-5.3-Flash-JANGH2-r2",
188
  "messages": [
189
  {
190
  "role": "user",
 
220
  }
221
  },
222
  "response": {
223
+ "id": "chatcmpl-51e2f985",
224
  "object": "chat.completion",
225
+ "created": 1790717262,
226
+ "model": "glm53-jangtq/GLM-5.3-Flash-JANGH2-r2",
227
  "choices": [
228
  {
229
  "index": 0,
 
231
  "role": "assistant",
232
  "tool_calls": [
233
  {
234
+ "id": "call_578fcc8c",
235
  "type": "function",
236
  "function": {
237
  "name": "lookup_code",
 
255
  }
256
  },
257
  "pass": true,
258
+ "seconds": 0.52547091699671
259
  },
260
  {
261
  "name": "tool p1 think=True choice=required",
262
  "request": {
263
+ "model": "glm53-jangtq/GLM-5.3-Flash-JANGH2-r2",
264
  "messages": [
265
  {
266
  "role": "user",
 
296
  }
297
  },
298
  "response": {
299
+ "id": "chatcmpl-9a22428f",
300
  "object": "chat.completion",
301
+ "created": 1790717264,
302
+ "model": "glm53-jangtq/GLM-5.3-Flash-JANGH2-r2",
303
  "choices": [
304
  {
305
  "index": 0,
 
307
  "role": "assistant",
308
  "tool_calls": [
309
  {
310
+ "id": "call_66cb738c",
311
  "type": "function",
312
  "function": {
313
  "name": "lookup_code",
 
315
  }
316
  }
317
  ],
318
+ "reasoning_content": "The user wants me to look up the code for label \"alpha\" using the lookup_code tool.",
319
  "content": null
320
  },
321
  "finish_reason": "tool_calls"
 
323
  ],
324
  "usage": {
325
  "prompt_tokens": 193,
326
+ "completion_tokens": 32,
327
+ "total_tokens": 225
 
 
 
 
328
  }
329
  },
330
  "pass": true,
331
+ "seconds": 2.2413089579786174
332
  },
333
  {
334
  "name": "tool p1 think=True choice=auto",
335
  "request": {
336
+ "model": "glm53-jangtq/GLM-5.3-Flash-JANGH2-r2",
337
  "messages": [
338
  {
339
  "role": "user",
 
369
  }
370
  },
371
  "response": {
372
+ "id": "chatcmpl-ecad5865",
373
  "object": "chat.completion",
374
+ "created": 1790717265,
375
+ "model": "glm53-jangtq/GLM-5.3-Flash-JANGH2-r2",
376
  "choices": [
377
  {
378
  "index": 0,
 
380
  "role": "assistant",
381
  "tool_calls": [
382
  {
383
+ "id": "call_73463e23",
384
  "type": "function",
385
  "function": {
386
  "name": "lookup_code",
 
388
  }
389
  }
390
  ],
391
+ "reasoning_content": "The user wants me to look up the code for label \"alpha\" using the lookup_code tool.",
392
  "content": null
393
  },
394
  "finish_reason": "tool_calls"
 
396
  ],
397
  "usage": {
398
  "prompt_tokens": 193,
399
+ "completion_tokens": 32,
400
+ "total_tokens": 225,
401
  "prompt_tokens_details": {
402
  "cached_tokens": 192,
403
  "cache_detail": "native-glm+disk"
 
405
  }
406
  },
407
  "pass": true,
408
+ "seconds": 1.2828379580168985
409
  },
410
  {
411
  "name": "tool p2 think=False choice=required",
412
  "request": {
413
+ "model": "glm53-jangtq/GLM-5.3-Flash-JANGH2-r2",
414
  "messages": [
415
  {
416
  "role": "user",
 
446
  }
447
  },
448
  "response": {
449
+ "id": "chatcmpl-13ea203c",
450
  "object": "chat.completion",
451
+ "created": 1790717267,
452
+ "model": "glm53-jangtq/GLM-5.3-Flash-JANGH2-r2",
453
  "choices": [
454
  {
455
  "index": 0,
 
457
  "role": "assistant",
458
  "tool_calls": [
459
  {
460
+ "id": "call_b288f0bc",
461
  "type": "function",
462
  "function": {
463
  "name": "lookup_code",
 
477
  }
478
  },
479
  "pass": true,
480
+ "seconds": 1.5650544160162099
481
  },
482
  {
483
  "name": "tool p2 think=False choice=auto",
484
  "request": {
485
+ "model": "glm53-jangtq/GLM-5.3-Flash-JANGH2-r2",
486
  "messages": [
487
  {
488
  "role": "user",
 
518
  }
519
  },
520
  "response": {
521
+ "id": "chatcmpl-634e136e",
522
  "object": "chat.completion",
523
+ "created": 1790717268,
524
+ "model": "glm53-jangtq/GLM-5.3-Flash-JANGH2-r2",
525
  "choices": [
526
  {
527
  "index": 0,
 
529
  "role": "assistant",
530
  "tool_calls": [
531
  {
532
+ "id": "call_4be64d92",
533
  "type": "function",
534
  "function": {
535
  "name": "lookup_code",
 
553
  }
554
  },
555
  "pass": true,
556
+ "seconds": 0.5339888340095058
557
  },
558
  {
559
  "name": "tool p2 think=True choice=required",
560
  "request": {
561
+ "model": "glm53-jangtq/GLM-5.3-Flash-JANGH2-r2",
562
  "messages": [
563
  {
564
  "role": "user",
 
594
  }
595
  },
596
  "response": {
597
+ "id": "chatcmpl-4e39f975",
598
  "object": "chat.completion",
599
+ "created": 1790717270,
600
+ "model": "glm53-jangtq/GLM-5.3-Flash-JANGH2-r2",
601
  "choices": [
602
  {
603
  "index": 0,
 
605
  "role": "assistant",
606
  "tool_calls": [
607
  {
608
+ "id": "call_7e5a7b23",
609
  "type": "function",
610
  "function": {
611
  "name": "lookup_code",
 
613
  }
614
  }
615
  ],
616
+ "reasoning_content": "The user wants me to look up the secret code for the label 'alpha'. I need to call the lookup_code tool with label 'alpha'.",
617
  "content": null
618
  },
619
  "finish_reason": "tool_calls"
 
621
  ],
622
  "usage": {
623
  "prompt_tokens": 200,
624
+ "completion_tokens": 41,
625
+ "total_tokens": 241
 
 
 
 
626
  }
627
  },
628
  "pass": true,
629
+ "seconds": 2.620424250024371
630
  },
631
  {
632
  "name": "tool p2 think=True choice=auto",
633
  "request": {
634
+ "model": "glm53-jangtq/GLM-5.3-Flash-JANGH2-r2",
635
  "messages": [
636
  {
637
  "role": "user",
 
667
  }
668
  },
669
  "response": {
670
+ "id": "chatcmpl-f8c9455d",
671
  "object": "chat.completion",
672
+ "created": 1790717272,
673
+ "model": "glm53-jangtq/GLM-5.3-Flash-JANGH2-r2",
674
  "choices": [
675
  {
676
  "index": 0,
 
678
  "role": "assistant",
679
  "tool_calls": [
680
  {
681
+ "id": "call_4af5a873",
682
  "type": "function",
683
  "function": {
684
  "name": "lookup_code",
 
686
  }
687
  }
688
  ],
689
+ "reasoning_content": "The user wants me to look up the secret code for the label 'alpha'. I need to call the lookup_code tool with label 'alpha'.",
690
  "content": null
691
  },
692
  "finish_reason": "tool_calls"
 
694
  ],
695
  "usage": {
696
  "prompt_tokens": 200,
697
+ "completion_tokens": 41,
698
+ "total_tokens": 241,
699
  "prompt_tokens_details": {
700
  "cached_tokens": 199,
701
  "cache_detail": "native-glm+disk"
 
703
  }
704
  },
705
  "pass": true,
706
+ "seconds": 1.602398417016957
707
  },
708
  {
709
  "name": "tool result round trip",
710
  "request": {
711
+ "model": "glm53-jangtq/GLM-5.3-Flash-JANGH2-r2",
712
  "messages": [
713
  {
714
  "role": "user",
 
718
  "role": "assistant",
719
  "tool_calls": [
720
  {
721
+ "id": "call_df3410cb",
722
  "type": "function",
723
  "function": {
724
  "name": "lookup_code",
 
730
  },
731
  {
732
  "role": "tool",
733
+ "tool_call_id": "call_df3410cb",
734
  "content": "{\"code\": \"BLUE-CEDAR\"}"
735
  }
736
  ],
 
762
  }
763
  },
764
  "response": {
765
+ "id": "chatcmpl-7a038768",
766
  "object": "chat.completion",
767
+ "created": 1790717273,
768
+ "model": "glm53-jangtq/GLM-5.3-Flash-JANGH2-r2",
769
  "choices": [
770
  {
771
  "index": 0,
 
792
  "context_exhaustion": null
793
  },
794
  "pass": true,
795
+ "seconds": 1.1969611250096932
796
  },
797
  {
798
  "name": "reasoning_effort=low",
799
  "request": {
800
+ "model": "glm53-jangtq/GLM-5.3-Flash-JANGH2-r2",
801
  "messages": [
802
  {
803
  "role": "user",
 
809
  "reasoning_effort": "low"
810
  },
811
  "response": {
812
+ "id": "chatcmpl-57775cd0",
813
  "object": "chat.completion",
814
+ "created": 1790717275,
815
+ "model": "glm53-jangtq/GLM-5.3-Flash-JANGH2-r2",
816
  "choices": [
817
  {
818
  "index": 0,
819
  "message": {
820
  "role": "assistant",
821
+ "content": "No. 221 = 13 \u00d7 17, so it's a composite number.",
822
  "tool_calls": null,
823
+ "reasoning_content": "221 = 13 \u00d7 17, not prime."
824
  },
825
  "logprobs": null,
826
  "finish_reason": "stop"
 
828
  ],
829
  "usage": {
830
  "prompt_tokens": 30,
831
+ "completion_tokens": 33,
832
+ "total_tokens": 63,
833
+ "prompt_tokens_details": null
 
 
 
834
  },
835
  "warnings": null,
836
  "context_exhaustion": null
837
  },
838
  "pass": true,
839
+ "seconds": 1.834579374990426
840
  },
841
  {
842
  "name": "reasoning_effort=high",
843
  "request": {
844
+ "model": "glm53-jangtq/GLM-5.3-Flash-JANGH2-r2",
845
  "messages": [
846
  {
847
  "role": "user",
 
853
  "reasoning_effort": "high"
854
  },
855
  "response": {
856
+ "id": "chatcmpl-92e454e0",
857
  "object": "chat.completion",
858
+ "created": 1790717279,
859
+ "model": "glm53-jangtq/GLM-5.3-Flash-JANGH2-r2",
860
  "choices": [
861
  {
862
  "index": 0,
863
  "message": {
864
  "role": "assistant",
865
+ "content": "No. 221 is not prime because it can be factored as 13 \u00d7 17.",
866
  "tool_calls": null,
867
+ "reasoning_content": "221 = 13 \u00d7 17, not prime."
868
  },
869
  "logprobs": null,
870
  "finish_reason": "stop"
 
872
  ],
873
  "usage": {
874
  "prompt_tokens": 30,
875
+ "completion_tokens": 35,
876
+ "total_tokens": 65,
877
+ "prompt_tokens_details": null
 
 
 
878
  },
879
  "warnings": null,
880
  "context_exhaustion": null
881
  },
882
  "pass": true,
883
+ "seconds": 4.108727833954617
884
  },
885
  {
886
  "name": "reasoning_effort=max",
887
  "request": {
888
+ "model": "glm53-jangtq/GLM-5.3-Flash-JANGH2-r2",
889
  "messages": [
890
  {
891
  "role": "user",
 
897
  "reasoning_effort": "max"
898
  },
899
  "response": {
900
+ "id": "chatcmpl-8e1e2785",
901
  "object": "chat.completion",
902
+ "created": 1790717282,
903
+ "model": "glm53-jangtq/GLM-5.3-Flash-JANGH2-r2",
904
  "choices": [
905
  {
906
  "index": 0,
907
  "message": {
908
  "role": "assistant",
909
+ "content": "No. 221 is not prime because it can be factored as 13 \u00d7 17.",
910
  "tool_calls": null,
911
+ "reasoning_content": "The user asks if 221 is prime. Let me check: 221 = 13 \u00d7 17. So it's not prime. Answer: No, because 221 = 13 \u00d7 17."
912
  },
913
  "logprobs": null,
914
  "finish_reason": "stop"
 
916
  ],
917
  "usage": {
918
  "prompt_tokens": 30,
919
+ "completion_tokens": 68,
920
+ "total_tokens": 98,
921
+ "prompt_tokens_details": null
 
 
 
922
  },
923
  "warnings": null,
924
  "context_exhaustion": null
925
  },
926
  "pass": true,
927
+ "seconds": 3.0803108330001123
928
  },
929
  {
930
  "name": "image colors/shapes",
931
  "request": {
932
+ "model": "glm53-jangtq/GLM-5.3-Flash-JANGH2-r2",
933
  "messages": [
934
  {
935
  "role": "user",
 
954
  }
955
  },
956
  "response": {
957
+ "id": "chatcmpl-a361db15",
958
  "object": "chat.completion",
959
+ "created": 1790717289,
960
+ "model": "glm53-jangtq/GLM-5.3-Flash-JANGH2-r2",
961
  "choices": [
962
  {
963
  "index": 0,
964
  "message": {
965
  "role": "assistant",
966
+ "content": "The user wants me to describe the visible colors and shapes in the image.\nLooking at the image:\n- There is a solid blue square in the center.\n- The background is white.\n- The shape is a square (or rectangle, but it looks square).\n- The color of the square is blue (specifically, pure blue or a bright blue, like #0000FF).\n- The background color is white.\n\nLet's write a simple, clear description of the colors and shapes.Based on the image, here is a description of the visible colors and shapes:\n\n* **Shapes:** There is a single **square** in the center of the image.\n* **Colors:** \n * The square is solid **blue** (specifically a bright, pure blue).\n * The background surrounding the square is **white**.",
967
  "tool_calls": null,
968
  "reasoning_content": null
969
  },
 
973
  ],
974
  "usage": {
975
  "prompt_tokens": 124,
976
+ "completion_tokens": 173,
977
+ "total_tokens": 297,
978
+ "prompt_tokens_details": null
 
 
 
979
  },
980
  "warnings": null,
981
  "context_exhaustion": null
982
  },
983
  "pass": true,
984
+ "seconds": 7.025973458017688
985
  },
986
  {
987
  "name": "video motion direction",
988
  "request": {
989
+ "model": "glm53-jangtq/GLM-5.3-Flash-JANGH2-r2",
990
  "messages": [
991
  {
992
  "role": "user",
 
1011
  }
1012
  },
1013
  "response": {
1014
+ "id": "chatcmpl-916ee30f",
1015
  "object": "chat.completion",
1016
+ "created": 1790717298,
1017
+ "model": "glm53-jangtq/GLM-5.3-Flash-JANGH2-r2",
1018
  "choices": [
1019
  {
1020
  "index": 0,
1021
  "message": {
1022
  "role": "assistant",
1023
+ "content": "The user wants to know the direction in which the colored square moves across the video.\nI will analyze the position of the square in each frame.\n- At 00:00, the square is on the left side of the screen.\n- At 00:02, the square is in the middle of the screen.\n- At 00:04, the square is still in the middle, but slightly to the right compared to 00:02.\n- At 00:06, the square is on the right side of the screen.\nThe square starts on the left and ends on the right. Therefore, it moves from left to right.Based on the video frames, the colored square starts on the left side of the screen at 00:00. By 00:02, it has moved to the center. It continues moving and is positioned on the right side of the screen by 00:06. Therefore, the square moves from left to right.",
1024
  "tool_calls": null,
1025
  "reasoning_content": null
1026
  },
 
1030
  ],
1031
  "usage": {
1032
  "prompt_tokens": 589,
1033
+ "completion_tokens": 197,
1034
+ "total_tokens": 786,
1035
+ "prompt_tokens_details": null
 
 
 
1036
  },
1037
  "warnings": null,
1038
  "context_exhaustion": null
1039
  },
1040
  "pass": true,
1041
+ "seconds": 9.237239582987968
1042
  }
1043
  ]
evaluation/kl_vs_fp8.json CHANGED
@@ -1,35 +1,35 @@
1
  {
2
  "positions": 15830,
3
- "median_kl": 0.032285298263584866,
4
- "mean_kl": 0.3512629079161063,
5
  "kl_percentiles": {
6
- "p50": 0.032285298263584866,
7
- "p75": 0.2576814605510518,
8
- "p90": 0.8918860567628951,
9
- "p95": 1.755830527218102,
10
- "p99": 4.525611581097773
11
  },
12
- "kl_max": 21.99042745227768,
13
- "top1_agreement_pct": 83.74605180037904,
14
- "ref_top1_in_top5_pct": 97.03095388502842,
15
- "ref_top1_in_top10_pct": 98.53442830069488,
16
  "flip_curve": [
17
  [
18
  0.0,
19
  0.7499999403953552,
20
- 0.533401325854156,
21
  1961
22
  ],
23
  [
24
  0.7499999403953552,
25
  1.7499999403953552,
26
- 0.32463622679377824,
27
  1993
28
  ],
29
  [
30
  1.7499999403953552,
31
  3.124999954365194,
32
- 0.17709384460141273,
33
  1982
34
  ],
35
  [
@@ -41,25 +41,25 @@
41
  [
42
  4.874999938998371,
43
  6.624999997293344,
44
- 0.07225871652349672,
45
  1979
46
  ],
47
  [
48
  6.624999997293344,
49
  8.500000116902811,
50
- 0.043478260869565216,
51
  1978
52
  ],
53
  [
54
  8.500000116902811,
55
  10.687500133566573,
56
- 0.020717534108135423,
57
  1979
58
  ],
59
  [
60
  10.687500133566573,
61
  23.8125,
62
- 0.007074279939363315,
63
  1979
64
  ]
65
  ],
 
1
  {
2
  "positions": 15830,
3
+ "median_kl": 0.030121881710082024,
4
+ "mean_kl": 0.37614552186788514,
5
  "kl_percentiles": {
6
+ "p50": 0.030121881710082024,
7
+ "p75": 0.2471745266366821,
8
+ "p90": 0.9753125174640521,
9
+ "p95": 1.9690169079681636,
10
+ "p99": 5.242001235392597
11
  },
12
+ "kl_max": 23.99713653286548,
13
+ "top1_agreement_pct": 83.90397978521794,
14
+ "ref_top1_in_top5_pct": 96.83512318382817,
15
+ "ref_top1_in_top10_pct": 98.27542640555906,
16
  "flip_curve": [
17
  [
18
  0.0,
19
  0.7499999403953552,
20
+ 0.5298317185109638,
21
  1961
22
  ],
23
  [
24
  0.7499999403953552,
25
  1.7499999403953552,
26
+ 0.30306071249372807,
27
  1993
28
  ],
29
  [
30
  1.7499999403953552,
31
  3.124999954365194,
32
+ 0.17507568113017155,
33
  1982
34
  ],
35
  [
 
41
  [
42
  4.874999938998371,
43
  6.624999997293344,
44
+ 0.07832238504295098,
45
  1979
46
  ],
47
  [
48
  6.624999997293344,
49
  8.500000116902811,
50
+ 0.051061678463094035,
51
  1978
52
  ],
53
  [
54
  8.500000116902811,
55
  10.687500133566573,
56
+ 0.021222839818089943,
57
  1979
58
  ],
59
  [
60
  10.687500133566573,
61
  23.8125,
62
+ 0.007579585649317837,
63
  1979
64
  ]
65
  ],
evaluation/native_transcripts_vs_bf16.json ADDED
@@ -0,0 +1,611 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "bundle": "GLM-5.3-Flash-JANGH2 (revision 2)",
3
+ "assistant": {
4
+ "positions": 60243,
5
+ "median_kl": 0.06881622225046158,
6
+ "mean_kl": 0.21784130269915855,
7
+ "p90": 0.5812361598014831,
8
+ "top1_pct": 82.5938283286025
9
+ },
10
+ "context": {
11
+ "positions": 7240,
12
+ "median_kl": 1.4476576447486877,
13
+ "mean_kl": 3.102558822176583,
14
+ "p90": 9.103206443786622,
15
+ "top1_pct": 45.027624309392266
16
+ },
17
+ "close": {
18
+ "n_all": 34,
19
+ "n_ref_ge_half": 20,
20
+ "argmax_correct": 14,
21
+ "p_q_ge_half": 12,
22
+ "median_p_q": 0.8495109677314758,
23
+ "min_p_q": 0.05205969512462616,
24
+ "each": [
25
+ {
26
+ "seq": "held_n00_agentic_low",
27
+ "pos": 407,
28
+ "p_ref": 0.9838244318962097,
29
+ "p_q": 0.2916111350059509,
30
+ "argmax_ok": false
31
+ },
32
+ {
33
+ "seq": "held_n00_agentic_low",
34
+ "pos": 588,
35
+ "p_ref": 0.7189128398895264,
36
+ "p_q": 0.3790413737297058,
37
+ "argmax_ok": false
38
+ },
39
+ {
40
+ "seq": "held_n00_agentic_low",
41
+ "pos": 965,
42
+ "p_ref": 0.8597419261932373,
43
+ "p_q": 0.299537718296051,
44
+ "argmax_ok": true
45
+ },
46
+ {
47
+ "seq": "held_n00_agentic_low",
48
+ "pos": 1116,
49
+ "p_ref": 0.25516748428344727,
50
+ "p_q": 0.20701055228710175,
51
+ "argmax_ok": false
52
+ },
53
+ {
54
+ "seq": "held_n00_agentic_low",
55
+ "pos": 1280,
56
+ "p_ref": 0.9985171556472778,
57
+ "p_q": 0.9975996017456055,
58
+ "argmax_ok": true
59
+ },
60
+ {
61
+ "seq": "held_n01_agentic_high",
62
+ "pos": 411,
63
+ "p_ref": 0.09345996379852295,
64
+ "p_q": 0.0008723821956664324,
65
+ "argmax_ok": false
66
+ },
67
+ {
68
+ "seq": "held_n01_agentic_high",
69
+ "pos": 481,
70
+ "p_ref": 0.9994451403617859,
71
+ "p_q": 0.9887097477912903,
72
+ "argmax_ok": true
73
+ },
74
+ {
75
+ "seq": "held_n01_agentic_high",
76
+ "pos": 818,
77
+ "p_ref": 0.05112185701727867,
78
+ "p_q": 0.01732826791703701,
79
+ "argmax_ok": false
80
+ },
81
+ {
82
+ "seq": "held_n01_agentic_high",
83
+ "pos": 999,
84
+ "p_ref": 0.010241878218948841,
85
+ "p_q": 0.016675179824233055,
86
+ "argmax_ok": false
87
+ },
88
+ {
89
+ "seq": "held_n01_agentic_high",
90
+ "pos": 1232,
91
+ "p_ref": 0.9731510877609253,
92
+ "p_q": 0.9661397933959961,
93
+ "argmax_ok": true
94
+ },
95
+ {
96
+ "seq": "held_n02_agentic_low",
97
+ "pos": 408,
98
+ "p_ref": 0.6796087026596069,
99
+ "p_q": 0.11788290739059448,
100
+ "argmax_ok": false
101
+ },
102
+ {
103
+ "seq": "held_n02_agentic_low",
104
+ "pos": 514,
105
+ "p_ref": 0.3617189824581146,
106
+ "p_q": 0.1208728477358818,
107
+ "argmax_ok": false
108
+ },
109
+ {
110
+ "seq": "held_n02_agentic_low",
111
+ "pos": 727,
112
+ "p_ref": 0.5420432090759277,
113
+ "p_q": 0.14295782148838043,
114
+ "argmax_ok": false
115
+ },
116
+ {
117
+ "seq": "held_n02_agentic_low",
118
+ "pos": 899,
119
+ "p_ref": 0.9765183329582214,
120
+ "p_q": 0.996669590473175,
121
+ "argmax_ok": true
122
+ },
123
+ {
124
+ "seq": "held_n03_agentic_high",
125
+ "pos": 410,
126
+ "p_ref": 0.5529681444168091,
127
+ "p_q": 0.05205969512462616,
128
+ "argmax_ok": false
129
+ },
130
+ {
131
+ "seq": "held_n03_agentic_high",
132
+ "pos": 488,
133
+ "p_ref": 0.743448793888092,
134
+ "p_q": 0.22527901828289032,
135
+ "argmax_ok": false
136
+ },
137
+ {
138
+ "seq": "held_n03_agentic_high",
139
+ "pos": 712,
140
+ "p_ref": 0.017172181978821754,
141
+ "p_q": 0.016983453184366226,
142
+ "argmax_ok": false
143
+ },
144
+ {
145
+ "seq": "held_n03_agentic_high",
146
+ "pos": 1193,
147
+ "p_ref": 0.013264749199151993,
148
+ "p_q": 0.02826460637152195,
149
+ "argmax_ok": false
150
+ },
151
+ {
152
+ "seq": "held_n03_agentic_high",
153
+ "pos": 1447,
154
+ "p_ref": 0.9964795112609863,
155
+ "p_q": 0.9916768074035645,
156
+ "argmax_ok": true
157
+ },
158
+ {
159
+ "seq": "held_n04_agentic_max",
160
+ "pos": 405,
161
+ "p_ref": 0.06360142678022385,
162
+ "p_q": 0.00022873509442433715,
163
+ "argmax_ok": false
164
+ },
165
+ {
166
+ "seq": "held_n04_agentic_max",
167
+ "pos": 491,
168
+ "p_ref": 0.09378442913293839,
169
+ "p_q": 0.06594368815422058,
170
+ "argmax_ok": false
171
+ },
172
+ {
173
+ "seq": "held_n04_agentic_max",
174
+ "pos": 717,
175
+ "p_ref": 0.022891120985150337,
176
+ "p_q": 0.007983765564858913,
177
+ "argmax_ok": false
178
+ },
179
+ {
180
+ "seq": "held_n04_agentic_max",
181
+ "pos": 973,
182
+ "p_ref": 0.06886352598667145,
183
+ "p_q": 0.010131229646503925,
184
+ "argmax_ok": false
185
+ },
186
+ {
187
+ "seq": "held_n04_agentic_max",
188
+ "pos": 1502,
189
+ "p_ref": 0.8409368395805359,
190
+ "p_q": 0.907599151134491,
191
+ "argmax_ok": true
192
+ },
193
+ {
194
+ "seq": "held_n05_agentic_low",
195
+ "pos": 408,
196
+ "p_ref": 0.9598525762557983,
197
+ "p_q": 0.46985504031181335,
198
+ "argmax_ok": true
199
+ },
200
+ {
201
+ "seq": "held_n05_agentic_low",
202
+ "pos": 478,
203
+ "p_ref": 0.9998569488525391,
204
+ "p_q": 0.9906918406486511,
205
+ "argmax_ok": true
206
+ },
207
+ {
208
+ "seq": "held_n05_agentic_low",
209
+ "pos": 815,
210
+ "p_ref": 0.025962302461266518,
211
+ "p_q": 0.02887396141886711,
212
+ "argmax_ok": false
213
+ },
214
+ {
215
+ "seq": "held_n05_agentic_low",
216
+ "pos": 982,
217
+ "p_ref": 0.27114608883857727,
218
+ "p_q": 0.19489246606826782,
219
+ "argmax_ok": false
220
+ },
221
+ {
222
+ "seq": "held_n05_agentic_low",
223
+ "pos": 1361,
224
+ "p_ref": 0.8944574594497681,
225
+ "p_q": 0.7914227843284607,
226
+ "argmax_ok": true
227
+ },
228
+ {
229
+ "seq": "held_28_design_low",
230
+ "pos": 8025,
231
+ "p_ref": 0.8218483328819275,
232
+ "p_q": 0.7522467970848083,
233
+ "argmax_ok": true
234
+ },
235
+ {
236
+ "seq": "held_32_coding_low",
237
+ "pos": 81,
238
+ "p_ref": 0.9979269504547119,
239
+ "p_q": 0.9863665699958801,
240
+ "argmax_ok": true
241
+ },
242
+ {
243
+ "seq": "held_34_agentic_high",
244
+ "pos": 370,
245
+ "p_ref": 0.9970080256462097,
246
+ "p_q": 0.9614063501358032,
247
+ "argmax_ok": true
248
+ },
249
+ {
250
+ "seq": "held_35_agentic_low",
251
+ "pos": 356,
252
+ "p_ref": 0.9999904632568359,
253
+ "p_q": 0.9999923706054688,
254
+ "argmax_ok": true
255
+ },
256
+ {
257
+ "seq": "held_36_general_high",
258
+ "pos": 177,
259
+ "p_ref": 0.3396732807159424,
260
+ "p_q": 0.36337098479270935,
261
+ "argmax_ok": false
262
+ }
263
+ ]
264
+ },
265
+ "tool_call": {
266
+ "n_all": 48,
267
+ "n_ref_ge_half": 41,
268
+ "argmax_correct": 40,
269
+ "p_q_ge_half": 40,
270
+ "median_p_q": 0.9998245239257812,
271
+ "min_p_q": 0.164503276348114,
272
+ "each": [
273
+ {
274
+ "seq": "held_n00_agentic_low",
275
+ "pos": 408,
276
+ "p_ref": 0.998909592628479,
277
+ "p_q": 0.9371992349624634,
278
+ "argmax_ok": true
279
+ },
280
+ {
281
+ "seq": "held_n00_agentic_low",
282
+ "pos": 421,
283
+ "p_ref": 0.4377860128879547,
284
+ "p_q": 0.24495118856430054,
285
+ "argmax_ok": false
286
+ },
287
+ {
288
+ "seq": "held_n00_agentic_low",
289
+ "pos": 589,
290
+ "p_ref": 0.9998664855957031,
291
+ "p_q": 0.9995080232620239,
292
+ "argmax_ok": true
293
+ },
294
+ {
295
+ "seq": "held_n00_agentic_low",
296
+ "pos": 616,
297
+ "p_ref": 0.5926499366760254,
298
+ "p_q": 0.164503276348114,
299
+ "argmax_ok": false
300
+ },
301
+ {
302
+ "seq": "held_n00_agentic_low",
303
+ "pos": 629,
304
+ "p_ref": 0.05340293049812317,
305
+ "p_q": 0.02033219300210476,
306
+ "argmax_ok": false
307
+ },
308
+ {
309
+ "seq": "held_n00_agentic_low",
310
+ "pos": 966,
311
+ "p_ref": 0.9999923706054688,
312
+ "p_q": 0.9998531341552734,
313
+ "argmax_ok": true
314
+ },
315
+ {
316
+ "seq": "held_n00_agentic_low",
317
+ "pos": 1117,
318
+ "p_ref": 0.9998550415039062,
319
+ "p_q": 0.9999332427978516,
320
+ "argmax_ok": true
321
+ },
322
+ {
323
+ "seq": "held_n00_agentic_low",
324
+ "pos": 1281,
325
+ "p_ref": 0.979510486125946,
326
+ "p_q": 0.9229016304016113,
327
+ "argmax_ok": true
328
+ },
329
+ {
330
+ "seq": "held_n01_agentic_high",
331
+ "pos": 412,
332
+ "p_ref": 0.0736844465136528,
333
+ "p_q": 0.057815179228782654,
334
+ "argmax_ok": false
335
+ },
336
+ {
337
+ "seq": "held_n01_agentic_high",
338
+ "pos": 482,
339
+ "p_ref": 0.9999847412109375,
340
+ "p_q": 0.9998798370361328,
341
+ "argmax_ok": true
342
+ },
343
+ {
344
+ "seq": "held_n01_agentic_high",
345
+ "pos": 495,
346
+ "p_ref": 0.9960519671440125,
347
+ "p_q": 0.9999198913574219,
348
+ "argmax_ok": true
349
+ },
350
+ {
351
+ "seq": "held_n01_agentic_high",
352
+ "pos": 508,
353
+ "p_ref": 0.9968274235725403,
354
+ "p_q": 0.9998416900634766,
355
+ "argmax_ok": true
356
+ },
357
+ {
358
+ "seq": "held_n01_agentic_high",
359
+ "pos": 840,
360
+ "p_ref": 0.9999809265136719,
361
+ "p_q": 0.9997101426124573,
362
+ "argmax_ok": true
363
+ },
364
+ {
365
+ "seq": "held_n01_agentic_high",
366
+ "pos": 1020,
367
+ "p_ref": 0.9999561309814453,
368
+ "p_q": 0.9999122619628906,
369
+ "argmax_ok": true
370
+ },
371
+ {
372
+ "seq": "held_n01_agentic_high",
373
+ "pos": 1251,
374
+ "p_ref": 1.0,
375
+ "p_q": 0.9999980926513672,
376
+ "argmax_ok": true
377
+ },
378
+ {
379
+ "seq": "held_n02_agentic_low",
380
+ "pos": 409,
381
+ "p_ref": 0.9621272683143616,
382
+ "p_q": 0.8799203038215637,
383
+ "argmax_ok": true
384
+ },
385
+ {
386
+ "seq": "held_n02_agentic_low",
387
+ "pos": 515,
388
+ "p_ref": 0.9767139554023743,
389
+ "p_q": 0.9765891432762146,
390
+ "argmax_ok": true
391
+ },
392
+ {
393
+ "seq": "held_n02_agentic_low",
394
+ "pos": 525,
395
+ "p_ref": 0.1191921979188919,
396
+ "p_q": 0.005220085848122835,
397
+ "argmax_ok": false
398
+ },
399
+ {
400
+ "seq": "held_n02_agentic_low",
401
+ "pos": 728,
402
+ "p_ref": 0.9975558519363403,
403
+ "p_q": 0.9956303238868713,
404
+ "argmax_ok": true
405
+ },
406
+ {
407
+ "seq": "held_n03_agentic_high",
408
+ "pos": 419,
409
+ "p_ref": 0.9999980926513672,
410
+ "p_q": 0.9999847412109375,
411
+ "argmax_ok": true
412
+ },
413
+ {
414
+ "seq": "held_n03_agentic_high",
415
+ "pos": 489,
416
+ "p_ref": 0.9332356452941895,
417
+ "p_q": 0.5371268391609192,
418
+ "argmax_ok": true
419
+ },
420
+ {
421
+ "seq": "held_n03_agentic_high",
422
+ "pos": 500,
423
+ "p_ref": 0.9999313354492188,
424
+ "p_q": 0.9999961853027344,
425
+ "argmax_ok": true
426
+ },
427
+ {
428
+ "seq": "held_n03_agentic_high",
429
+ "pos": 744,
430
+ "p_ref": 0.999969482421875,
431
+ "p_q": 0.9998912811279297,
432
+ "argmax_ok": true
433
+ },
434
+ {
435
+ "seq": "held_n03_agentic_high",
436
+ "pos": 757,
437
+ "p_ref": 0.9999675750732422,
438
+ "p_q": 0.9999961853027344,
439
+ "argmax_ok": true
440
+ },
441
+ {
442
+ "seq": "held_n03_agentic_high",
443
+ "pos": 770,
444
+ "p_ref": 0.9924128651618958,
445
+ "p_q": 0.999079167842865,
446
+ "argmax_ok": true
447
+ },
448
+ {
449
+ "seq": "held_n03_agentic_high",
450
+ "pos": 783,
451
+ "p_ref": 0.9525591731071472,
452
+ "p_q": 0.9840852618217468,
453
+ "argmax_ok": true
454
+ },
455
+ {
456
+ "seq": "held_n03_agentic_high",
457
+ "pos": 1219,
458
+ "p_ref": 0.9999675750732422,
459
+ "p_q": 0.9999217987060547,
460
+ "argmax_ok": true
461
+ },
462
+ {
463
+ "seq": "held_n03_agentic_high",
464
+ "pos": 1448,
465
+ "p_ref": 0.06318042427301407,
466
+ "p_q": 0.1402663141489029,
467
+ "argmax_ok": false
468
+ },
469
+ {
470
+ "seq": "held_n03_agentic_high",
471
+ "pos": 1461,
472
+ "p_ref": 0.9982962608337402,
473
+ "p_q": 0.9998245239257812,
474
+ "argmax_ok": true
475
+ },
476
+ {
477
+ "seq": "held_n03_agentic_high",
478
+ "pos": 1474,
479
+ "p_ref": 0.9889983534812927,
480
+ "p_q": 0.9989610314369202,
481
+ "argmax_ok": true
482
+ },
483
+ {
484
+ "seq": "held_n04_agentic_max",
485
+ "pos": 422,
486
+ "p_ref": 0.999969482421875,
487
+ "p_q": 0.9999313354492188,
488
+ "argmax_ok": true
489
+ },
490
+ {
491
+ "seq": "held_n04_agentic_max",
492
+ "pos": 492,
493
+ "p_ref": 0.8776340484619141,
494
+ "p_q": 0.928158700466156,
495
+ "argmax_ok": true
496
+ },
497
+ {
498
+ "seq": "held_n04_agentic_max",
499
+ "pos": 505,
500
+ "p_ref": 0.946465015411377,
501
+ "p_q": 0.9958600997924805,
502
+ "argmax_ok": true
503
+ },
504
+ {
505
+ "seq": "held_n04_agentic_max",
506
+ "pos": 748,
507
+ "p_ref": 0.9997082352638245,
508
+ "p_q": 0.9995251893997192,
509
+ "argmax_ok": true
510
+ },
511
+ {
512
+ "seq": "held_n04_agentic_max",
513
+ "pos": 761,
514
+ "p_ref": 0.9995595216751099,
515
+ "p_q": 0.9998397827148438,
516
+ "argmax_ok": true
517
+ },
518
+ {
519
+ "seq": "held_n04_agentic_max",
520
+ "pos": 994,
521
+ "p_ref": 0.9999771118164062,
522
+ "p_q": 0.9999160766601562,
523
+ "argmax_ok": true
524
+ },
525
+ {
526
+ "seq": "held_n04_agentic_max",
527
+ "pos": 1009,
528
+ "p_ref": 0.9999809265136719,
529
+ "p_q": 0.9999961853027344,
530
+ "argmax_ok": true
531
+ },
532
+ {
533
+ "seq": "held_n04_agentic_max",
534
+ "pos": 1022,
535
+ "p_ref": 0.9768145084381104,
536
+ "p_q": 0.9889700412750244,
537
+ "argmax_ok": true
538
+ },
539
+ {
540
+ "seq": "held_n04_agentic_max",
541
+ "pos": 1035,
542
+ "p_ref": 0.320767343044281,
543
+ "p_q": 0.622435450553894,
544
+ "argmax_ok": true
545
+ },
546
+ {
547
+ "seq": "held_n04_agentic_max",
548
+ "pos": 1532,
549
+ "p_ref": 0.9999961853027344,
550
+ "p_q": 0.9999923706054688,
551
+ "argmax_ok": true
552
+ },
553
+ {
554
+ "seq": "held_n05_agentic_low",
555
+ "pos": 409,
556
+ "p_ref": 0.9866337776184082,
557
+ "p_q": 0.9518871903419495,
558
+ "argmax_ok": true
559
+ },
560
+ {
561
+ "seq": "held_n05_agentic_low",
562
+ "pos": 479,
563
+ "p_ref": 0.9999980926513672,
564
+ "p_q": 0.9999504089355469,
565
+ "argmax_ok": true
566
+ },
567
+ {
568
+ "seq": "held_n05_agentic_low",
569
+ "pos": 490,
570
+ "p_ref": 0.9999637603759766,
571
+ "p_q": 0.9999866485595703,
572
+ "argmax_ok": true
573
+ },
574
+ {
575
+ "seq": "held_n05_agentic_low",
576
+ "pos": 503,
577
+ "p_ref": 0.9525591731071472,
578
+ "p_q": 0.7057181596755981,
579
+ "argmax_ok": true
580
+ },
581
+ {
582
+ "seq": "held_n05_agentic_low",
583
+ "pos": 832,
584
+ "p_ref": 0.9999046325683594,
585
+ "p_q": 0.9996529221534729,
586
+ "argmax_ok": true
587
+ },
588
+ {
589
+ "seq": "held_n05_agentic_low",
590
+ "pos": 1030,
591
+ "p_ref": 0.0414167195558548,
592
+ "p_q": 0.1559947282075882,
593
+ "argmax_ok": false
594
+ },
595
+ {
596
+ "seq": "held_34_agentic_high",
597
+ "pos": 381,
598
+ "p_ref": 0.9999942779541016,
599
+ "p_q": 0.9999866485595703,
600
+ "argmax_ok": true
601
+ },
602
+ {
603
+ "seq": "held_35_agentic_low",
604
+ "pos": 357,
605
+ "p_ref": 0.9996204972267151,
606
+ "p_q": 0.9953587651252747,
607
+ "argmax_ok": true
608
+ }
609
+ ]
610
+ }
611
+ }
evaluation/scale_correction_report.json ADDED
@@ -0,0 +1,884 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "model.layers.5.mlp.switch_mlp.gate_proj": {
3
+ "bits": 3,
4
+ "gain_mean": 1.0358974933624268,
5
+ "gain_min": 1.0235940217971802,
6
+ "gain_max": 1.0484700202941895,
7
+ "clipped_rows": 0
8
+ },
9
+ "model.layers.4.mlp.switch_mlp.down_proj": {
10
+ "bits": 3,
11
+ "gain_mean": 1.0359731912612915,
12
+ "gain_min": 0.9749687314033508,
13
+ "gain_max": 1.102686882019043,
14
+ "clipped_rows": 0
15
+ },
16
+ "model.layers.4.mlp.switch_mlp.gate_proj": {
17
+ "bits": 2,
18
+ "gain_mean": 1.1320784091949463,
19
+ "gain_min": 1.094233751296997,
20
+ "gain_max": 1.1563719511032104,
21
+ "clipped_rows": 0
22
+ },
23
+ "model.layers.3.mlp.switch_mlp.gate_proj": {
24
+ "bits": 2,
25
+ "gain_mean": 1.13102388381958,
26
+ "gain_min": 1.0871330499649048,
27
+ "gain_max": 1.1566838026046753,
28
+ "clipped_rows": 0
29
+ },
30
+ "model.layers.4.mlp.switch_mlp.up_proj": {
31
+ "bits": 2,
32
+ "gain_mean": 1.1322484016418457,
33
+ "gain_min": 1.102554202079773,
34
+ "gain_max": 1.1613961458206177,
35
+ "clipped_rows": 0
36
+ },
37
+ "model.layers.3.mlp.switch_mlp.down_proj": {
38
+ "bits": 3,
39
+ "gain_mean": 1.0366361141204834,
40
+ "gain_min": 0.8598466515541077,
41
+ "gain_max": 1.227429747581482,
42
+ "clipped_rows": 0
43
+ },
44
+ "model.layers.3.mlp.switch_mlp.up_proj": {
45
+ "bits": 2,
46
+ "gain_mean": 1.131264328956604,
47
+ "gain_min": 1.091144323348999,
48
+ "gain_max": 1.1569643020629883,
49
+ "clipped_rows": 0
50
+ },
51
+ "model.layers.7.mlp.switch_mlp.up_proj": {
52
+ "bits": 2,
53
+ "gain_mean": 1.1323515176773071,
54
+ "gain_min": 1.1068731546401978,
55
+ "gain_max": 1.1577699184417725,
56
+ "clipped_rows": 0
57
+ },
58
+ "model.layers.7.mlp.switch_mlp.gate_proj": {
59
+ "bits": 2,
60
+ "gain_mean": 1.1321457624435425,
61
+ "gain_min": 1.101946473121643,
62
+ "gain_max": 1.1580171585083008,
63
+ "clipped_rows": 0
64
+ },
65
+ "model.layers.6.mlp.switch_mlp.up_proj": {
66
+ "bits": 3,
67
+ "gain_mean": 1.0359277725219727,
68
+ "gain_min": 1.0228333473205566,
69
+ "gain_max": 1.0464410781860352,
70
+ "clipped_rows": 0
71
+ },
72
+ "model.layers.6.mlp.switch_mlp.gate_proj": {
73
+ "bits": 3,
74
+ "gain_mean": 1.0358885526657104,
75
+ "gain_min": 1.0254722833633423,
76
+ "gain_max": 1.0464026927947998,
77
+ "clipped_rows": 0
78
+ },
79
+ "model.layers.5.mlp.switch_mlp.up_proj": {
80
+ "bits": 3,
81
+ "gain_mean": 1.0359306335449219,
82
+ "gain_min": 1.0231802463531494,
83
+ "gain_max": 1.0470621585845947,
84
+ "clipped_rows": 0
85
+ },
86
+ "model.layers.6.mlp.switch_mlp.down_proj": {
87
+ "bits": 3,
88
+ "gain_mean": 1.0359668731689453,
89
+ "gain_min": 0.8116584420204163,
90
+ "gain_max": 1.8735743761062622,
91
+ "clipped_rows": 0
92
+ },
93
+ "model.layers.5.mlp.switch_mlp.down_proj": {
94
+ "bits": 3,
95
+ "gain_mean": 1.035942792892456,
96
+ "gain_min": 0.9455461502075195,
97
+ "gain_max": 1.111367106437683,
98
+ "clipped_rows": 0
99
+ },
100
+ "model.layers.9.mlp.switch_mlp.gate_proj": {
101
+ "bits": 2,
102
+ "gain_mean": 1.131778359413147,
103
+ "gain_min": 1.092459797859192,
104
+ "gain_max": 1.1572468280792236,
105
+ "clipped_rows": 0
106
+ },
107
+ "model.layers.9.mlp.switch_mlp.down_proj": {
108
+ "bits": 3,
109
+ "gain_mean": 1.0359973907470703,
110
+ "gain_min": 0.8795009255409241,
111
+ "gain_max": 1.1949326992034912,
112
+ "clipped_rows": 0
113
+ },
114
+ "model.layers.8.mlp.switch_mlp.up_proj": {
115
+ "bits": 2,
116
+ "gain_mean": 1.131937861442566,
117
+ "gain_min": 1.1053141355514526,
118
+ "gain_max": 1.1571930646896362,
119
+ "clipped_rows": 0
120
+ },
121
+ "model.layers.8.mlp.switch_mlp.down_proj": {
122
+ "bits": 3,
123
+ "gain_mean": 1.0359998941421509,
124
+ "gain_min": 0.8455630540847778,
125
+ "gain_max": 1.2587127685546875,
126
+ "clipped_rows": 0
127
+ },
128
+ "model.layers.7.mlp.switch_mlp.down_proj": {
129
+ "bits": 3,
130
+ "gain_mean": 1.035959005355835,
131
+ "gain_min": 0.8108063340187073,
132
+ "gain_max": 1.8600410223007202,
133
+ "clipped_rows": 0
134
+ },
135
+ "model.layers.9.mlp.switch_mlp.up_proj": {
136
+ "bits": 2,
137
+ "gain_mean": 1.1320146322250366,
138
+ "gain_min": 1.0987129211425781,
139
+ "gain_max": 1.1547508239746094,
140
+ "clipped_rows": 0
141
+ },
142
+ "model.layers.8.mlp.switch_mlp.gate_proj": {
143
+ "bits": 2,
144
+ "gain_mean": 1.131662368774414,
145
+ "gain_min": 1.0991698503494263,
146
+ "gain_max": 1.165192723274231,
147
+ "clipped_rows": 0
148
+ },
149
+ "model.layers.10.mlp.switch_mlp.gate_proj": {
150
+ "bits": 2,
151
+ "gain_mean": 1.1318868398666382,
152
+ "gain_min": 1.0872851610183716,
153
+ "gain_max": 1.1598132848739624,
154
+ "clipped_rows": 0
155
+ },
156
+ "model.layers.12.mlp.switch_mlp.up_proj": {
157
+ "bits": 2,
158
+ "gain_mean": 1.1309666633605957,
159
+ "gain_min": 1.0873336791992188,
160
+ "gain_max": 1.1552129983901978,
161
+ "clipped_rows": 0
162
+ },
163
+ "model.layers.12.mlp.switch_mlp.gate_proj": {
164
+ "bits": 2,
165
+ "gain_mean": 1.1305294036865234,
166
+ "gain_min": 1.096046805381775,
167
+ "gain_max": 1.1550722122192383,
168
+ "clipped_rows": 0
169
+ },
170
+ "model.layers.12.mlp.switch_mlp.down_proj": {
171
+ "bits": 3,
172
+ "gain_mean": 1.035934329032898,
173
+ "gain_min": 0.9730667471885681,
174
+ "gain_max": 1.0752508640289307,
175
+ "clipped_rows": 0
176
+ },
177
+ "model.layers.11.mlp.switch_mlp.gate_proj": {
178
+ "bits": 2,
179
+ "gain_mean": 1.1317908763885498,
180
+ "gain_min": 1.1028674840927124,
181
+ "gain_max": 1.1583253145217896,
182
+ "clipped_rows": 0
183
+ },
184
+ "model.layers.11.mlp.switch_mlp.up_proj": {
185
+ "bits": 2,
186
+ "gain_mean": 1.1320468187332153,
187
+ "gain_min": 1.1003355979919434,
188
+ "gain_max": 1.1554360389709473,
189
+ "clipped_rows": 0
190
+ },
191
+ "model.layers.11.mlp.switch_mlp.down_proj": {
192
+ "bits": 3,
193
+ "gain_mean": 1.0360161066055298,
194
+ "gain_min": 0.9441719055175781,
195
+ "gain_max": 1.1149201393127441,
196
+ "clipped_rows": 0
197
+ },
198
+ "model.layers.10.mlp.switch_mlp.up_proj": {
199
+ "bits": 2,
200
+ "gain_mean": 1.1321293115615845,
201
+ "gain_min": 1.0834853649139404,
202
+ "gain_max": 1.159075140953064,
203
+ "clipped_rows": 0
204
+ },
205
+ "model.layers.10.mlp.switch_mlp.down_proj": {
206
+ "bits": 3,
207
+ "gain_mean": 1.0360313653945923,
208
+ "gain_min": 0.742192268371582,
209
+ "gain_max": 2.0,
210
+ "clipped_rows": 1
211
+ },
212
+ "model.layers.15.mlp.switch_mlp.up_proj": {
213
+ "bits": 2,
214
+ "gain_mean": 1.1318641901016235,
215
+ "gain_min": 1.096354603767395,
216
+ "gain_max": 1.156887173652649,
217
+ "clipped_rows": 0
218
+ },
219
+ "model.layers.15.mlp.switch_mlp.gate_proj": {
220
+ "bits": 2,
221
+ "gain_mean": 1.1314008235931396,
222
+ "gain_min": 1.0951993465423584,
223
+ "gain_max": 1.1566227674484253,
224
+ "clipped_rows": 0
225
+ },
226
+ "model.layers.14.mlp.switch_mlp.down_proj": {
227
+ "bits": 3,
228
+ "gain_mean": 1.0359647274017334,
229
+ "gain_min": 0.8767577409744263,
230
+ "gain_max": 1.3292709589004517,
231
+ "clipped_rows": 0
232
+ },
233
+ "model.layers.13.mlp.switch_mlp.down_proj": {
234
+ "bits": 3,
235
+ "gain_mean": 1.0359293222427368,
236
+ "gain_min": 0.9637039303779602,
237
+ "gain_max": 1.1065748929977417,
238
+ "clipped_rows": 0
239
+ },
240
+ "model.layers.13.mlp.switch_mlp.up_proj": {
241
+ "bits": 2,
242
+ "gain_mean": 1.130875587463379,
243
+ "gain_min": 1.0995205640792847,
244
+ "gain_max": 1.1557599306106567,
245
+ "clipped_rows": 0
246
+ },
247
+ "model.layers.14.mlp.switch_mlp.up_proj": {
248
+ "bits": 2,
249
+ "gain_mean": 1.1310826539993286,
250
+ "gain_min": 1.0912266969680786,
251
+ "gain_max": 1.1566081047058105,
252
+ "clipped_rows": 0
253
+ },
254
+ "model.layers.14.mlp.switch_mlp.gate_proj": {
255
+ "bits": 2,
256
+ "gain_mean": 1.1305228471755981,
257
+ "gain_min": 1.094700574874878,
258
+ "gain_max": 1.1584848165512085,
259
+ "clipped_rows": 0
260
+ },
261
+ "model.layers.13.mlp.switch_mlp.gate_proj": {
262
+ "bits": 2,
263
+ "gain_mean": 1.130236268043518,
264
+ "gain_min": 1.096887469291687,
265
+ "gain_max": 1.1549816131591797,
266
+ "clipped_rows": 0
267
+ },
268
+ "model.layers.18.mlp.switch_mlp.gate_proj": {
269
+ "bits": 2,
270
+ "gain_mean": 1.130363941192627,
271
+ "gain_min": 1.0984755754470825,
272
+ "gain_max": 1.1547136306762695,
273
+ "clipped_rows": 0
274
+ },
275
+ "model.layers.17.mlp.switch_mlp.up_proj": {
276
+ "bits": 2,
277
+ "gain_mean": 1.1308420896530151,
278
+ "gain_min": 1.1008402109146118,
279
+ "gain_max": 1.1586247682571411,
280
+ "clipped_rows": 0
281
+ },
282
+ "model.layers.17.mlp.switch_mlp.gate_proj": {
283
+ "bits": 2,
284
+ "gain_mean": 1.1303300857543945,
285
+ "gain_min": 1.0988088846206665,
286
+ "gain_max": 1.1560529470443726,
287
+ "clipped_rows": 0
288
+ },
289
+ "model.layers.16.mlp.switch_mlp.up_proj": {
290
+ "bits": 2,
291
+ "gain_mean": 1.131499171257019,
292
+ "gain_min": 1.095237135887146,
293
+ "gain_max": 1.1566752195358276,
294
+ "clipped_rows": 0
295
+ },
296
+ "model.layers.17.mlp.switch_mlp.down_proj": {
297
+ "bits": 3,
298
+ "gain_mean": 1.0358628034591675,
299
+ "gain_min": 0.974286675453186,
300
+ "gain_max": 1.075804352760315,
301
+ "clipped_rows": 0
302
+ },
303
+ "model.layers.16.mlp.switch_mlp.gate_proj": {
304
+ "bits": 2,
305
+ "gain_mean": 1.1310490369796753,
306
+ "gain_min": 1.0940485000610352,
307
+ "gain_max": 1.156827688217163,
308
+ "clipped_rows": 0
309
+ },
310
+ "model.layers.16.mlp.switch_mlp.down_proj": {
311
+ "bits": 3,
312
+ "gain_mean": 1.035904884338379,
313
+ "gain_min": 0.8209829926490784,
314
+ "gain_max": 1.2795387506484985,
315
+ "clipped_rows": 0
316
+ },
317
+ "model.layers.15.mlp.switch_mlp.down_proj": {
318
+ "bits": 3,
319
+ "gain_mean": 1.0359450578689575,
320
+ "gain_min": 0.9606896042823792,
321
+ "gain_max": 1.1033520698547363,
322
+ "clipped_rows": 0
323
+ },
324
+ "model.layers.20.mlp.switch_mlp.gate_proj": {
325
+ "bits": 2,
326
+ "gain_mean": 1.1304670572280884,
327
+ "gain_min": 1.101560115814209,
328
+ "gain_max": 1.1563524007797241,
329
+ "clipped_rows": 0
330
+ },
331
+ "model.layers.19.mlp.switch_mlp.gate_proj": {
332
+ "bits": 2,
333
+ "gain_mean": 1.1307493448257446,
334
+ "gain_min": 1.1030123233795166,
335
+ "gain_max": 1.1563241481781006,
336
+ "clipped_rows": 0
337
+ },
338
+ "model.layers.20.mlp.switch_mlp.down_proj": {
339
+ "bits": 3,
340
+ "gain_mean": 1.0358507633209229,
341
+ "gain_min": 0.9554993510246277,
342
+ "gain_max": 1.1591477394104004,
343
+ "clipped_rows": 0
344
+ },
345
+ "model.layers.19.mlp.switch_mlp.up_proj": {
346
+ "bits": 2,
347
+ "gain_mean": 1.131243348121643,
348
+ "gain_min": 1.098380446434021,
349
+ "gain_max": 1.1555417776107788,
350
+ "clipped_rows": 0
351
+ },
352
+ "model.layers.19.mlp.switch_mlp.down_proj": {
353
+ "bits": 3,
354
+ "gain_mean": 1.0358480215072632,
355
+ "gain_min": 0.936430811882019,
356
+ "gain_max": 1.169289231300354,
357
+ "clipped_rows": 0
358
+ },
359
+ "model.layers.18.mlp.switch_mlp.up_proj": {
360
+ "bits": 2,
361
+ "gain_mean": 1.1309564113616943,
362
+ "gain_min": 1.0962542295455933,
363
+ "gain_max": 1.1592429876327515,
364
+ "clipped_rows": 0
365
+ },
366
+ "model.layers.18.mlp.switch_mlp.down_proj": {
367
+ "bits": 3,
368
+ "gain_mean": 1.0359363555908203,
369
+ "gain_min": 0.9591401815414429,
370
+ "gain_max": 1.2295012474060059,
371
+ "clipped_rows": 0
372
+ },
373
+ "model.layers.20.mlp.switch_mlp.up_proj": {
374
+ "bits": 2,
375
+ "gain_mean": 1.1309484243392944,
376
+ "gain_min": 1.1015936136245728,
377
+ "gain_max": 1.1559892892837524,
378
+ "clipped_rows": 0
379
+ },
380
+ "model.layers.23.mlp.switch_mlp.up_proj": {
381
+ "bits": 2,
382
+ "gain_mean": 1.1312966346740723,
383
+ "gain_min": 1.1022075414657593,
384
+ "gain_max": 1.1570459604263306,
385
+ "clipped_rows": 0
386
+ },
387
+ "model.layers.22.mlp.switch_mlp.up_proj": {
388
+ "bits": 2,
389
+ "gain_mean": 1.1311873197555542,
390
+ "gain_min": 1.0913565158843994,
391
+ "gain_max": 1.1556634902954102,
392
+ "clipped_rows": 0
393
+ },
394
+ "model.layers.21.mlp.switch_mlp.down_proj": {
395
+ "bits": 3,
396
+ "gain_mean": 1.0358507633209229,
397
+ "gain_min": 0.9610126614570618,
398
+ "gain_max": 1.1081207990646362,
399
+ "clipped_rows": 0
400
+ },
401
+ "model.layers.21.mlp.switch_mlp.up_proj": {
402
+ "bits": 2,
403
+ "gain_mean": 1.1307977437973022,
404
+ "gain_min": 1.1011309623718262,
405
+ "gain_max": 1.158044695854187,
406
+ "clipped_rows": 0
407
+ },
408
+ "model.layers.22.mlp.switch_mlp.down_proj": {
409
+ "bits": 3,
410
+ "gain_mean": 1.0359110832214355,
411
+ "gain_min": 0.9614159464836121,
412
+ "gain_max": 1.108801007270813,
413
+ "clipped_rows": 0
414
+ },
415
+ "model.layers.23.mlp.switch_mlp.gate_proj": {
416
+ "bits": 2,
417
+ "gain_mean": 1.130903959274292,
418
+ "gain_min": 1.098525881767273,
419
+ "gain_max": 1.1536517143249512,
420
+ "clipped_rows": 0
421
+ },
422
+ "model.layers.22.mlp.switch_mlp.gate_proj": {
423
+ "bits": 2,
424
+ "gain_mean": 1.1308188438415527,
425
+ "gain_min": 1.0931047201156616,
426
+ "gain_max": 1.1602932214736938,
427
+ "clipped_rows": 0
428
+ },
429
+ "model.layers.21.mlp.switch_mlp.gate_proj": {
430
+ "bits": 2,
431
+ "gain_mean": 1.1303154230117798,
432
+ "gain_min": 1.0941146612167358,
433
+ "gain_max": 1.1572155952453613,
434
+ "clipped_rows": 0
435
+ },
436
+ "model.layers.26.mlp.switch_mlp.up_proj": {
437
+ "bits": 2,
438
+ "gain_mean": 1.1317927837371826,
439
+ "gain_min": 1.1038202047348022,
440
+ "gain_max": 1.1552788019180298,
441
+ "clipped_rows": 0
442
+ },
443
+ "model.layers.26.mlp.switch_mlp.gate_proj": {
444
+ "bits": 2,
445
+ "gain_mean": 1.1314159631729126,
446
+ "gain_min": 1.1018409729003906,
447
+ "gain_max": 1.1556651592254639,
448
+ "clipped_rows": 0
449
+ },
450
+ "model.layers.25.mlp.switch_mlp.up_proj": {
451
+ "bits": 2,
452
+ "gain_mean": 1.1314588785171509,
453
+ "gain_min": 1.095633625984192,
454
+ "gain_max": 1.1569453477859497,
455
+ "clipped_rows": 0
456
+ },
457
+ "model.layers.25.mlp.switch_mlp.gate_proj": {
458
+ "bits": 2,
459
+ "gain_mean": 1.131105661392212,
460
+ "gain_min": 1.0994430780410767,
461
+ "gain_max": 1.155840277671814,
462
+ "clipped_rows": 0
463
+ },
464
+ "model.layers.23.mlp.switch_mlp.down_proj": {
465
+ "bits": 3,
466
+ "gain_mean": 1.0358344316482544,
467
+ "gain_min": 0.9758607745170593,
468
+ "gain_max": 1.082773208618164,
469
+ "clipped_rows": 0
470
+ },
471
+ "model.layers.24.mlp.switch_mlp.up_proj": {
472
+ "bits": 2,
473
+ "gain_mean": 1.1313421726226807,
474
+ "gain_min": 1.098559856414795,
475
+ "gain_max": 1.1558512449264526,
476
+ "clipped_rows": 0
477
+ },
478
+ "model.layers.24.mlp.switch_mlp.down_proj": {
479
+ "bits": 3,
480
+ "gain_mean": 1.0359560251235962,
481
+ "gain_min": 0.8873920440673828,
482
+ "gain_max": 1.2387827634811401,
483
+ "clipped_rows": 0
484
+ },
485
+ "model.layers.25.mlp.switch_mlp.down_proj": {
486
+ "bits": 2,
487
+ "gain_mean": 1.1333688497543335,
488
+ "gain_min": 0.9214738607406616,
489
+ "gain_max": 1.6676524877548218,
490
+ "clipped_rows": 0
491
+ },
492
+ "model.layers.24.mlp.switch_mlp.gate_proj": {
493
+ "bits": 2,
494
+ "gain_mean": 1.130895733833313,
495
+ "gain_min": 1.0999696254730225,
496
+ "gain_max": 1.1560968160629272,
497
+ "clipped_rows": 0
498
+ },
499
+ "model.layers.29.mlp.switch_mlp.gate_proj": {
500
+ "bits": 2,
501
+ "gain_mean": 1.1317389011383057,
502
+ "gain_min": 1.104701042175293,
503
+ "gain_max": 1.1558979749679565,
504
+ "clipped_rows": 0
505
+ },
506
+ "model.layers.28.mlp.switch_mlp.up_proj": {
507
+ "bits": 2,
508
+ "gain_mean": 1.1318950653076172,
509
+ "gain_min": 1.1020443439483643,
510
+ "gain_max": 1.1561068296432495,
511
+ "clipped_rows": 0
512
+ },
513
+ "model.layers.28.mlp.switch_mlp.down_proj": {
514
+ "bits": 3,
515
+ "gain_mean": 1.0359867811203003,
516
+ "gain_min": 0.9191966652870178,
517
+ "gain_max": 1.1963213682174683,
518
+ "clipped_rows": 0
519
+ },
520
+ "model.layers.26.mlp.switch_mlp.down_proj": {
521
+ "bits": 3,
522
+ "gain_mean": 1.0359680652618408,
523
+ "gain_min": 0.9952130913734436,
524
+ "gain_max": 1.0864323377609253,
525
+ "clipped_rows": 0
526
+ },
527
+ "model.layers.27.mlp.switch_mlp.up_proj": {
528
+ "bits": 2,
529
+ "gain_mean": 1.1318367719650269,
530
+ "gain_min": 1.1058578491210938,
531
+ "gain_max": 1.1558457612991333,
532
+ "clipped_rows": 0
533
+ },
534
+ "model.layers.27.mlp.switch_mlp.gate_proj": {
535
+ "bits": 2,
536
+ "gain_mean": 1.1314808130264282,
537
+ "gain_min": 1.0993852615356445,
538
+ "gain_max": 1.1585602760314941,
539
+ "clipped_rows": 0
540
+ },
541
+ "model.layers.27.mlp.switch_mlp.down_proj": {
542
+ "bits": 3,
543
+ "gain_mean": 1.0359846353530884,
544
+ "gain_min": 1.011968970298767,
545
+ "gain_max": 1.0661500692367554,
546
+ "clipped_rows": 0
547
+ },
548
+ "model.layers.28.mlp.switch_mlp.gate_proj": {
549
+ "bits": 2,
550
+ "gain_mean": 1.1314857006072998,
551
+ "gain_min": 1.1030755043029785,
552
+ "gain_max": 1.1593385934829712,
553
+ "clipped_rows": 0
554
+ },
555
+ "model.layers.31.mlp.switch_mlp.gate_proj": {
556
+ "bits": 2,
557
+ "gain_mean": 1.1318455934524536,
558
+ "gain_min": 1.1004778146743774,
559
+ "gain_max": 1.1557732820510864,
560
+ "clipped_rows": 0
561
+ },
562
+ "model.layers.30.mlp.switch_mlp.up_proj": {
563
+ "bits": 2,
564
+ "gain_mean": 1.132112979888916,
565
+ "gain_min": 1.0989017486572266,
566
+ "gain_max": 1.1568360328674316,
567
+ "clipped_rows": 0
568
+ },
569
+ "model.layers.31.mlp.switch_mlp.down_proj": {
570
+ "bits": 3,
571
+ "gain_mean": 1.035974383354187,
572
+ "gain_min": 1.001680612564087,
573
+ "gain_max": 1.0833569765090942,
574
+ "clipped_rows": 0
575
+ },
576
+ "model.layers.31.mlp.switch_mlp.up_proj": {
577
+ "bits": 2,
578
+ "gain_mean": 1.1322683095932007,
579
+ "gain_min": 1.102881908416748,
580
+ "gain_max": 1.1562021970748901,
581
+ "clipped_rows": 0
582
+ },
583
+ "model.layers.30.mlp.switch_mlp.gate_proj": {
584
+ "bits": 2,
585
+ "gain_mean": 1.1316521167755127,
586
+ "gain_min": 1.1043182611465454,
587
+ "gain_max": 1.1569664478302002,
588
+ "clipped_rows": 0
589
+ },
590
+ "model.layers.30.mlp.switch_mlp.down_proj": {
591
+ "bits": 3,
592
+ "gain_mean": 1.0359965562820435,
593
+ "gain_min": 0.9217982888221741,
594
+ "gain_max": 1.144686222076416,
595
+ "clipped_rows": 0
596
+ },
597
+ "model.layers.29.mlp.switch_mlp.down_proj": {
598
+ "bits": 3,
599
+ "gain_mean": 1.0359797477722168,
600
+ "gain_min": 0.8722316026687622,
601
+ "gain_max": 1.2713004350662231,
602
+ "clipped_rows": 0
603
+ },
604
+ "model.layers.29.mlp.switch_mlp.up_proj": {
605
+ "bits": 2,
606
+ "gain_mean": 1.1321526765823364,
607
+ "gain_min": 1.1009868383407593,
608
+ "gain_max": 1.1550335884094238,
609
+ "clipped_rows": 0
610
+ },
611
+ "model.layers.34.mlp.switch_mlp.gate_proj": {
612
+ "bits": 3,
613
+ "gain_mean": 1.0357739925384521,
614
+ "gain_min": 1.0255333185195923,
615
+ "gain_max": 1.0465835332870483,
616
+ "clipped_rows": 0
617
+ },
618
+ "model.layers.32.mlp.switch_mlp.gate_proj": {
619
+ "bits": 3,
620
+ "gain_mean": 1.0358080863952637,
621
+ "gain_min": 1.0243128538131714,
622
+ "gain_max": 1.0503891706466675,
623
+ "clipped_rows": 0
624
+ },
625
+ "model.layers.33.mlp.switch_mlp.gate_proj": {
626
+ "bits": 2,
627
+ "gain_mean": 1.1318081617355347,
628
+ "gain_min": 1.1027112007141113,
629
+ "gain_max": 1.1574516296386719,
630
+ "clipped_rows": 0
631
+ },
632
+ "model.layers.33.mlp.switch_mlp.up_proj": {
633
+ "bits": 2,
634
+ "gain_mean": 1.1323261260986328,
635
+ "gain_min": 1.0996854305267334,
636
+ "gain_max": 1.1566252708435059,
637
+ "clipped_rows": 0
638
+ },
639
+ "model.layers.33.mlp.switch_mlp.down_proj": {
640
+ "bits": 3,
641
+ "gain_mean": 1.0359666347503662,
642
+ "gain_min": 0.9731428027153015,
643
+ "gain_max": 1.086214542388916,
644
+ "clipped_rows": 0
645
+ },
646
+ "model.layers.32.mlp.switch_mlp.up_proj": {
647
+ "bits": 3,
648
+ "gain_mean": 1.0359243154525757,
649
+ "gain_min": 1.0204983949661255,
650
+ "gain_max": 1.047611951828003,
651
+ "clipped_rows": 0
652
+ },
653
+ "model.layers.32.mlp.switch_mlp.down_proj": {
654
+ "bits": 3,
655
+ "gain_mean": 1.0359684228897095,
656
+ "gain_min": 0.9444416165351868,
657
+ "gain_max": 1.1281341314315796,
658
+ "clipped_rows": 0
659
+ },
660
+ "model.layers.34.mlp.switch_mlp.down_proj": {
661
+ "bits": 3,
662
+ "gain_mean": 1.035961389541626,
663
+ "gain_min": 0.9313869476318359,
664
+ "gain_max": 1.2453007698059082,
665
+ "clipped_rows": 0
666
+ },
667
+ "model.layers.35.mlp.switch_mlp.gate_proj": {
668
+ "bits": 3,
669
+ "gain_mean": 1.0358253717422485,
670
+ "gain_min": 1.0231105089187622,
671
+ "gain_max": 1.0465391874313354,
672
+ "clipped_rows": 0
673
+ },
674
+ "model.layers.35.mlp.switch_mlp.up_proj": {
675
+ "bits": 3,
676
+ "gain_mean": 1.0359302759170532,
677
+ "gain_min": 1.0223897695541382,
678
+ "gain_max": 1.0478571653366089,
679
+ "clipped_rows": 0
680
+ },
681
+ "model.layers.36.mlp.switch_mlp.gate_proj": {
682
+ "bits": 3,
683
+ "gain_mean": 1.0358045101165771,
684
+ "gain_min": 1.024606466293335,
685
+ "gain_max": 1.0457581281661987,
686
+ "clipped_rows": 0
687
+ },
688
+ "model.layers.35.mlp.switch_mlp.down_proj": {
689
+ "bits": 3,
690
+ "gain_mean": 1.035946249961853,
691
+ "gain_min": 0.9329993724822998,
692
+ "gain_max": 1.1259303092956543,
693
+ "clipped_rows": 0
694
+ },
695
+ "model.layers.34.mlp.switch_mlp.up_proj": {
696
+ "bits": 3,
697
+ "gain_mean": 1.035892128944397,
698
+ "gain_min": 1.0251798629760742,
699
+ "gain_max": 1.0464106798171997,
700
+ "clipped_rows": 0
701
+ },
702
+ "model.layers.38.mlp.switch_mlp.up_proj": {
703
+ "bits": 2,
704
+ "gain_mean": 1.13246750831604,
705
+ "gain_min": 1.0795625448226929,
706
+ "gain_max": 1.1585625410079956,
707
+ "clipped_rows": 0
708
+ },
709
+ "model.layers.38.mlp.switch_mlp.down_proj": {
710
+ "bits": 3,
711
+ "gain_mean": 1.0359662771224976,
712
+ "gain_min": 0.96891188621521,
713
+ "gain_max": 1.126165509223938,
714
+ "clipped_rows": 0
715
+ },
716
+ "model.layers.37.mlp.switch_mlp.up_proj": {
717
+ "bits": 2,
718
+ "gain_mean": 1.1324902772903442,
719
+ "gain_min": 1.1040645837783813,
720
+ "gain_max": 1.1573982238769531,
721
+ "clipped_rows": 0
722
+ },
723
+ "model.layers.36.mlp.switch_mlp.up_proj": {
724
+ "bits": 3,
725
+ "gain_mean": 1.0359249114990234,
726
+ "gain_min": 1.0177196264266968,
727
+ "gain_max": 1.0461890697479248,
728
+ "clipped_rows": 0
729
+ },
730
+ "model.layers.37.mlp.switch_mlp.gate_proj": {
731
+ "bits": 2,
732
+ "gain_mean": 1.1318656206130981,
733
+ "gain_min": 1.1047329902648926,
734
+ "gain_max": 1.1603649854660034,
735
+ "clipped_rows": 0
736
+ },
737
+ "model.layers.37.mlp.switch_mlp.down_proj": {
738
+ "bits": 3,
739
+ "gain_mean": 1.0359452962875366,
740
+ "gain_min": 0.8920966386795044,
741
+ "gain_max": 1.386698842048645,
742
+ "clipped_rows": 0
743
+ },
744
+ "model.layers.38.mlp.switch_mlp.gate_proj": {
745
+ "bits": 2,
746
+ "gain_mean": 1.1318423748016357,
747
+ "gain_min": 1.0995676517486572,
748
+ "gain_max": 1.1569567918777466,
749
+ "clipped_rows": 0
750
+ },
751
+ "model.layers.36.mlp.switch_mlp.down_proj": {
752
+ "bits": 3,
753
+ "gain_mean": 1.0359268188476562,
754
+ "gain_min": 0.9906195998191833,
755
+ "gain_max": 1.0865460634231567,
756
+ "clipped_rows": 0
757
+ },
758
+ "model.layers.41.mlp.switch_mlp.up_proj": {
759
+ "bits": 2,
760
+ "gain_mean": 1.13222336769104,
761
+ "gain_min": 1.0979671478271484,
762
+ "gain_max": 1.1555283069610596,
763
+ "clipped_rows": 0
764
+ },
765
+ "model.layers.41.mlp.switch_mlp.gate_proj": {
766
+ "bits": 2,
767
+ "gain_mean": 1.1313034296035767,
768
+ "gain_min": 1.0957648754119873,
769
+ "gain_max": 1.1562687158584595,
770
+ "clipped_rows": 0
771
+ },
772
+ "model.layers.40.mlp.switch_mlp.gate_proj": {
773
+ "bits": 2,
774
+ "gain_mean": 1.1316348314285278,
775
+ "gain_min": 1.0999752283096313,
776
+ "gain_max": 1.155924677848816,
777
+ "clipped_rows": 0
778
+ },
779
+ "model.layers.40.mlp.switch_mlp.up_proj": {
780
+ "bits": 2,
781
+ "gain_mean": 1.1323906183242798,
782
+ "gain_min": 1.1025410890579224,
783
+ "gain_max": 1.162571668624878,
784
+ "clipped_rows": 0
785
+ },
786
+ "model.layers.40.mlp.switch_mlp.down_proj": {
787
+ "bits": 3,
788
+ "gain_mean": 1.0359456539154053,
789
+ "gain_min": 0.984533965587616,
790
+ "gain_max": 1.0796700716018677,
791
+ "clipped_rows": 0
792
+ },
793
+ "model.layers.39.mlp.switch_mlp.gate_proj": {
794
+ "bits": 2,
795
+ "gain_mean": 1.131907343864441,
796
+ "gain_min": 1.0998762845993042,
797
+ "gain_max": 1.159606695175171,
798
+ "clipped_rows": 0
799
+ },
800
+ "model.layers.39.mlp.switch_mlp.up_proj": {
801
+ "bits": 2,
802
+ "gain_mean": 1.1325165033340454,
803
+ "gain_min": 1.1038825511932373,
804
+ "gain_max": 1.156624674797058,
805
+ "clipped_rows": 0
806
+ },
807
+ "model.layers.39.mlp.switch_mlp.down_proj": {
808
+ "bits": 3,
809
+ "gain_mean": 1.035938024520874,
810
+ "gain_min": 1.0003584623336792,
811
+ "gain_max": 1.0605535507202148,
812
+ "clipped_rows": 0
813
+ },
814
+ "model.layers.43.mlp.switch_mlp.up_proj": {
815
+ "bits": 2,
816
+ "gain_mean": 1.1320743560791016,
817
+ "gain_min": 1.0954220294952393,
818
+ "gain_max": 1.1580873727798462,
819
+ "clipped_rows": 0
820
+ },
821
+ "model.layers.43.mlp.switch_mlp.gate_proj": {
822
+ "bits": 2,
823
+ "gain_mean": 1.1313408613204956,
824
+ "gain_min": 1.0914368629455566,
825
+ "gain_max": 1.1624529361724854,
826
+ "clipped_rows": 0
827
+ },
828
+ "model.layers.43.mlp.switch_mlp.down_proj": {
829
+ "bits": 3,
830
+ "gain_mean": 1.036327600479126,
831
+ "gain_min": 0.8422228097915649,
832
+ "gain_max": 1.166949987411499,
833
+ "clipped_rows": 0
834
+ },
835
+ "model.layers.42.mlp.switch_mlp.gate_proj": {
836
+ "bits": 2,
837
+ "gain_mean": 1.1311098337173462,
838
+ "gain_min": 1.0928794145584106,
839
+ "gain_max": 1.162437915802002,
840
+ "clipped_rows": 0
841
+ },
842
+ "model.layers.44.mlp.switch_mlp.gate_proj": {
843
+ "bits": 2,
844
+ "gain_mean": 1.1314083337783813,
845
+ "gain_min": 0.8074365258216858,
846
+ "gain_max": 1.160056233406067,
847
+ "clipped_rows": 0
848
+ },
849
+ "model.layers.42.mlp.switch_mlp.up_proj": {
850
+ "bits": 2,
851
+ "gain_mean": 1.1320817470550537,
852
+ "gain_min": 1.0927314758300781,
853
+ "gain_max": 1.1606314182281494,
854
+ "clipped_rows": 0
855
+ },
856
+ "model.layers.42.mlp.switch_mlp.down_proj": {
857
+ "bits": 3,
858
+ "gain_mean": 1.0363696813583374,
859
+ "gain_min": 0.9331578016281128,
860
+ "gain_max": 1.1196359395980835,
861
+ "clipped_rows": 0
862
+ },
863
+ "model.layers.41.mlp.switch_mlp.down_proj": {
864
+ "bits": 3,
865
+ "gain_mean": 1.0359832048416138,
866
+ "gain_min": 0.9525346159934998,
867
+ "gain_max": 1.1272246837615967,
868
+ "clipped_rows": 0
869
+ },
870
+ "model.layers.44.mlp.switch_mlp.down_proj": {
871
+ "bits": 3,
872
+ "gain_mean": 1.0364047288894653,
873
+ "gain_min": 0.8899533748626709,
874
+ "gain_max": 1.2003225088119507,
875
+ "clipped_rows": 0
876
+ },
877
+ "model.layers.44.mlp.switch_mlp.up_proj": {
878
+ "bits": 2,
879
+ "gain_mean": 1.1317473649978638,
880
+ "gain_min": 0.7754930257797241,
881
+ "gain_max": 1.4476484060287476,
882
+ "clipped_rows": 0
883
+ }
884
+ }
evaluation/served_reasoning_matrix.json ADDED
@@ -0,0 +1,786 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "case": "math",
4
+ "effort": "low",
5
+ "seed": 1,
6
+ "clear_thinking": true,
7
+ "note": "",
8
+ "completion_tokens": 63,
9
+ "prompt_tokens": 20,
10
+ "reasoning_chars": 31,
11
+ "content_chars": 83,
12
+ "finish": "stop",
13
+ "reasoning_closed": true,
14
+ "seconds": 3.5
15
+ },
16
+ {
17
+ "case": "math",
18
+ "effort": "low",
19
+ "seed": 2,
20
+ "clear_thinking": true,
21
+ "note": "",
22
+ "completion_tokens": 46,
23
+ "prompt_tokens": 20,
24
+ "reasoning_chars": 13,
25
+ "content_chars": 98,
26
+ "finish": "stop",
27
+ "reasoning_closed": true,
28
+ "seconds": 1.8
29
+ },
30
+ {
31
+ "case": "math",
32
+ "effort": "low",
33
+ "seed": 3,
34
+ "clear_thinking": true,
35
+ "note": "",
36
+ "completion_tokens": 58,
37
+ "prompt_tokens": 20,
38
+ "reasoning_chars": 38,
39
+ "content_chars": 83,
40
+ "finish": "stop",
41
+ "reasoning_closed": true,
42
+ "seconds": 2.2
43
+ },
44
+ {
45
+ "case": "intervals",
46
+ "effort": "low",
47
+ "seed": 1,
48
+ "clear_thinking": true,
49
+ "note": "",
50
+ "completion_tokens": 569,
51
+ "prompt_tokens": 26,
52
+ "reasoning_chars": 12,
53
+ "content_chars": 2305,
54
+ "finish": "stop",
55
+ "reasoning_closed": true,
56
+ "seconds": 21.3
57
+ },
58
+ {
59
+ "case": "intervals",
60
+ "effort": "low",
61
+ "seed": 2,
62
+ "clear_thinking": true,
63
+ "note": "",
64
+ "completion_tokens": 542,
65
+ "prompt_tokens": 26,
66
+ "reasoning_chars": 0,
67
+ "content_chars": 2171,
68
+ "finish": "stop",
69
+ "reasoning_closed": true,
70
+ "seconds": 20.0
71
+ },
72
+ {
73
+ "case": "intervals",
74
+ "effort": "low",
75
+ "seed": 3,
76
+ "clear_thinking": true,
77
+ "note": "",
78
+ "completion_tokens": 513,
79
+ "prompt_tokens": 26,
80
+ "reasoning_chars": 0,
81
+ "content_chars": 1946,
82
+ "finish": "stop",
83
+ "reasoning_closed": true,
84
+ "seconds": 18.9
85
+ },
86
+ {
87
+ "case": "ssh",
88
+ "effort": "low",
89
+ "seed": 1,
90
+ "clear_thinking": true,
91
+ "note": "",
92
+ "completion_tokens": 748,
93
+ "prompt_tokens": 31,
94
+ "reasoning_chars": 289,
95
+ "content_chars": 2515,
96
+ "finish": "stop",
97
+ "reasoning_closed": true,
98
+ "seconds": 28.1
99
+ },
100
+ {
101
+ "case": "ssh",
102
+ "effort": "low",
103
+ "seed": 2,
104
+ "clear_thinking": true,
105
+ "note": "",
106
+ "completion_tokens": 692,
107
+ "prompt_tokens": 31,
108
+ "reasoning_chars": 0,
109
+ "content_chars": 2599,
110
+ "finish": "stop",
111
+ "reasoning_closed": true,
112
+ "seconds": 25.9
113
+ },
114
+ {
115
+ "case": "ssh",
116
+ "effort": "low",
117
+ "seed": 3,
118
+ "clear_thinking": true,
119
+ "note": "",
120
+ "completion_tokens": 809,
121
+ "prompt_tokens": 31,
122
+ "reasoning_chars": 0,
123
+ "content_chars": 3088,
124
+ "finish": "stop",
125
+ "reasoning_closed": true,
126
+ "seconds": 30.7
127
+ },
128
+ {
129
+ "case": "garbled_adlab",
130
+ "effort": "default(max)",
131
+ "seed": 1,
132
+ "clear_thinking": true,
133
+ "note": "",
134
+ "completion_tokens": 788,
135
+ "prompt_tokens": 34,
136
+ "reasoning_chars": 2372,
137
+ "content_chars": 719,
138
+ "finish": "stop",
139
+ "reasoning_closed": true,
140
+ "seconds": 30.5
141
+ },
142
+ {
143
+ "case": "garbled_adlab",
144
+ "effort": "default(max)",
145
+ "seed": 2,
146
+ "clear_thinking": true,
147
+ "note": "",
148
+ "completion_tokens": 1070,
149
+ "prompt_tokens": 34,
150
+ "reasoning_chars": 3382,
151
+ "content_chars": 754,
152
+ "finish": "stop",
153
+ "reasoning_closed": true,
154
+ "seconds": 40.6
155
+ },
156
+ {
157
+ "case": "garbled_adlab",
158
+ "effort": "default(max)",
159
+ "seed": 3,
160
+ "clear_thinking": true,
161
+ "note": "",
162
+ "completion_tokens": 922,
163
+ "prompt_tokens": 34,
164
+ "reasoning_chars": 2677,
165
+ "content_chars": 660,
166
+ "finish": "stop",
167
+ "reasoning_closed": true,
168
+ "seconds": 34.9
169
+ },
170
+ {
171
+ "case": "garbled_adlab",
172
+ "effort": "default(max)",
173
+ "seed": 4,
174
+ "clear_thinking": true,
175
+ "note": "",
176
+ "completion_tokens": 1015,
177
+ "prompt_tokens": 34,
178
+ "reasoning_chars": 3070,
179
+ "content_chars": 722,
180
+ "finish": "stop",
181
+ "reasoning_closed": true,
182
+ "seconds": 38.5
183
+ },
184
+ {
185
+ "case": "garbled_adlab",
186
+ "effort": "default(max)",
187
+ "seed": 5,
188
+ "clear_thinking": true,
189
+ "note": "",
190
+ "completion_tokens": 886,
191
+ "prompt_tokens": 34,
192
+ "reasoning_chars": 2651,
193
+ "content_chars": 652,
194
+ "finish": "stop",
195
+ "reasoning_closed": true,
196
+ "seconds": 33.6
197
+ },
198
+ {
199
+ "case": "garbled_adlab",
200
+ "effort": "default(max)",
201
+ "seed": 6,
202
+ "clear_thinking": true,
203
+ "note": "",
204
+ "completion_tokens": 935,
205
+ "prompt_tokens": 34,
206
+ "reasoning_chars": 2990,
207
+ "content_chars": 522,
208
+ "finish": "stop",
209
+ "reasoning_closed": true,
210
+ "seconds": 35.5
211
+ },
212
+ {
213
+ "case": "isp",
214
+ "effort": "low",
215
+ "seed": 1,
216
+ "clear_thinking": true,
217
+ "note": "",
218
+ "completion_tokens": 16000,
219
+ "prompt_tokens": 70,
220
+ "reasoning_chars": 27388,
221
+ "content_chars": 28306,
222
+ "finish": "length",
223
+ "reasoning_closed": true,
224
+ "seconds": 655.0
225
+ },
226
+ {
227
+ "case": "isp",
228
+ "effort": "low",
229
+ "seed": 2,
230
+ "clear_thinking": true,
231
+ "note": "",
232
+ "completion_tokens": 12327,
233
+ "prompt_tokens": 70,
234
+ "reasoning_chars": 18302,
235
+ "content_chars": 25215,
236
+ "finish": "stop",
237
+ "reasoning_closed": true,
238
+ "seconds": 509.0
239
+ },
240
+ {
241
+ "case": "isp",
242
+ "effort": "low",
243
+ "seed": 3,
244
+ "clear_thinking": true,
245
+ "note": "",
246
+ "completion_tokens": 16000,
247
+ "prompt_tokens": 70,
248
+ "reasoning_chars": 30260,
249
+ "content_chars": 23723,
250
+ "finish": "length",
251
+ "reasoning_closed": true,
252
+ "seconds": 666.6
253
+ },
254
+ {
255
+ "case": "isp",
256
+ "effort": "max",
257
+ "seed": 1,
258
+ "clear_thinking": true,
259
+ "note": "",
260
+ "completion_tokens": 16000,
261
+ "prompt_tokens": 70,
262
+ "reasoning_chars": 57743,
263
+ "content_chars": 0,
264
+ "finish": "length",
265
+ "reasoning_closed": false,
266
+ "seconds": 666.4
267
+ },
268
+ {
269
+ "case": "isp",
270
+ "effort": "max",
271
+ "seed": 2,
272
+ "clear_thinking": true,
273
+ "note": "",
274
+ "completion_tokens": 16000,
275
+ "prompt_tokens": 70,
276
+ "reasoning_chars": 57337,
277
+ "content_chars": 0,
278
+ "finish": "length",
279
+ "reasoning_closed": false,
280
+ "seconds": 663.0
281
+ },
282
+ {
283
+ "case": "multiturn",
284
+ "effort": "low",
285
+ "seed": 11,
286
+ "clear_thinking": true,
287
+ "note": "turn 1 OK",
288
+ "completion_tokens": 84,
289
+ "prompt_tokens": 44,
290
+ "reasoning_chars": 57,
291
+ "content_chars": 94,
292
+ "finish": "stop",
293
+ "reasoning_closed": true,
294
+ "seconds": 4.0
295
+ },
296
+ {
297
+ "case": "multiturn",
298
+ "effort": "low",
299
+ "seed": 12,
300
+ "clear_thinking": true,
301
+ "note": "turn 2 OK",
302
+ "completion_tokens": 62,
303
+ "prompt_tokens": 98,
304
+ "reasoning_chars": 44,
305
+ "content_chars": 101,
306
+ "finish": "stop",
307
+ "reasoning_closed": true,
308
+ "seconds": 3.2
309
+ },
310
+ {
311
+ "case": "multiturn",
312
+ "effort": "low",
313
+ "seed": 13,
314
+ "clear_thinking": true,
315
+ "note": "turn 3 OK",
316
+ "completion_tokens": 203,
317
+ "prompt_tokens": 155,
318
+ "reasoning_chars": 107,
319
+ "content_chars": 402,
320
+ "finish": "stop",
321
+ "reasoning_closed": true,
322
+ "seconds": 8.7
323
+ },
324
+ {
325
+ "case": "multiturn",
326
+ "effort": "low",
327
+ "seed": 14,
328
+ "clear_thinking": true,
329
+ "note": "turn 4 OK",
330
+ "completion_tokens": 161,
331
+ "prompt_tokens": 321,
332
+ "reasoning_chars": 51,
333
+ "content_chars": 479,
334
+ "finish": "stop",
335
+ "reasoning_closed": true,
336
+ "seconds": 7.3
337
+ },
338
+ {
339
+ "case": "multiturn",
340
+ "effort": "low",
341
+ "seed": 11,
342
+ "clear_thinking": false,
343
+ "note": "turn 1 OK",
344
+ "completion_tokens": 84,
345
+ "prompt_tokens": 44,
346
+ "reasoning_chars": 57,
347
+ "content_chars": 94,
348
+ "finish": "stop",
349
+ "reasoning_closed": true,
350
+ "seconds": 3.3
351
+ },
352
+ {
353
+ "case": "multiturn",
354
+ "effort": "low",
355
+ "seed": 12,
356
+ "clear_thinking": false,
357
+ "note": "turn 2 OK",
358
+ "completion_tokens": 62,
359
+ "prompt_tokens": 134,
360
+ "reasoning_chars": 44,
361
+ "content_chars": 101,
362
+ "finish": "stop",
363
+ "reasoning_closed": true,
364
+ "seconds": 3.2
365
+ },
366
+ {
367
+ "case": "multiturn",
368
+ "effort": "low",
369
+ "seed": 13,
370
+ "clear_thinking": false,
371
+ "note": "turn 3 OK",
372
+ "completion_tokens": 181,
373
+ "prompt_tokens": 213,
374
+ "reasoning_chars": 88,
375
+ "content_chars": 423,
376
+ "finish": "stop",
377
+ "reasoning_closed": true,
378
+ "seconds": 7.8
379
+ },
380
+ {
381
+ "case": "multiturn",
382
+ "effort": "low",
383
+ "seed": 14,
384
+ "clear_thinking": false,
385
+ "note": "turn 4 OK",
386
+ "completion_tokens": 131,
387
+ "prompt_tokens": 414,
388
+ "reasoning_chars": 51,
389
+ "content_chars": 354,
390
+ "finish": "stop",
391
+ "reasoning_closed": true,
392
+ "seconds": 6.3
393
+ },
394
+ {
395
+ "case": "multiturn",
396
+ "effort": "max",
397
+ "seed": 11,
398
+ "clear_thinking": true,
399
+ "note": "turn 1 OK",
400
+ "completion_tokens": 185,
401
+ "prompt_tokens": 44,
402
+ "reasoning_chars": 361,
403
+ "content_chars": 104,
404
+ "finish": "stop",
405
+ "reasoning_closed": true,
406
+ "seconds": 7.9
407
+ },
408
+ {
409
+ "case": "multiturn",
410
+ "effort": "max",
411
+ "seed": 12,
412
+ "clear_thinking": true,
413
+ "note": "turn 2 OK",
414
+ "completion_tokens": 93,
415
+ "prompt_tokens": 99,
416
+ "reasoning_chars": 108,
417
+ "content_chars": 122,
418
+ "finish": "stop",
419
+ "reasoning_closed": true,
420
+ "seconds": 4.4
421
+ },
422
+ {
423
+ "case": "multiturn",
424
+ "effort": "max",
425
+ "seed": 13,
426
+ "clear_thinking": true,
427
+ "note": "turn 3 OK",
428
+ "completion_tokens": 360,
429
+ "prompt_tokens": 161,
430
+ "reasoning_chars": 353,
431
+ "content_chars": 549,
432
+ "finish": "stop",
433
+ "reasoning_closed": true,
434
+ "seconds": 14.8
435
+ },
436
+ {
437
+ "case": "multiturn",
438
+ "effort": "max",
439
+ "seed": 14,
440
+ "clear_thinking": true,
441
+ "note": "turn 4 OK",
442
+ "completion_tokens": 455,
443
+ "prompt_tokens": 380,
444
+ "reasoning_chars": 832,
445
+ "content_chars": 530,
446
+ "finish": "stop",
447
+ "reasoning_closed": true,
448
+ "seconds": 18.9
449
+ },
450
+ {
451
+ "case": "multiturn",
452
+ "effort": "max",
453
+ "seed": 11,
454
+ "clear_thinking": false,
455
+ "note": "turn 1 OK",
456
+ "completion_tokens": 185,
457
+ "prompt_tokens": 44,
458
+ "reasoning_chars": 361,
459
+ "content_chars": 104,
460
+ "finish": "stop",
461
+ "reasoning_closed": true,
462
+ "seconds": 7.2
463
+ },
464
+ {
465
+ "case": "multiturn",
466
+ "effort": "max",
467
+ "seed": 12,
468
+ "clear_thinking": false,
469
+ "note": "turn 2 OK",
470
+ "completion_tokens": 116,
471
+ "prompt_tokens": 235,
472
+ "reasoning_chars": 199,
473
+ "content_chars": 121,
474
+ "finish": "stop",
475
+ "reasoning_closed": true,
476
+ "seconds": 5.6
477
+ },
478
+ {
479
+ "case": "multiturn",
480
+ "effort": "max",
481
+ "seed": 13,
482
+ "clear_thinking": false,
483
+ "note": "turn 3 OK",
484
+ "completion_tokens": 304,
485
+ "prompt_tokens": 368,
486
+ "reasoning_chars": 397,
487
+ "content_chars": 441,
488
+ "finish": "stop",
489
+ "reasoning_closed": true,
490
+ "seconds": 12.8
491
+ },
492
+ {
493
+ "case": "multiturn",
494
+ "effort": "max",
495
+ "seed": 14,
496
+ "clear_thinking": false,
497
+ "note": "turn 4 OK",
498
+ "completion_tokens": 306,
499
+ "prompt_tokens": 692,
500
+ "reasoning_chars": 491,
501
+ "content_chars": 410,
502
+ "finish": "stop",
503
+ "reasoning_closed": true,
504
+ "seconds": 13.4
505
+ },
506
+ {
507
+ "case": "math",
508
+ "effort": "high",
509
+ "seed": 1,
510
+ "clear_thinking": true,
511
+ "note": "",
512
+ "completion_tokens": 69,
513
+ "prompt_tokens": 20,
514
+ "reasoning_chars": 39,
515
+ "content_chars": 96,
516
+ "finish": "stop",
517
+ "reasoning_closed": true,
518
+ "seconds": 3.7
519
+ },
520
+ {
521
+ "case": "math",
522
+ "effort": "max",
523
+ "seed": 1,
524
+ "clear_thinking": true,
525
+ "note": "",
526
+ "completion_tokens": 158,
527
+ "prompt_tokens": 20,
528
+ "reasoning_chars": 267,
529
+ "content_chars": 81,
530
+ "finish": "stop",
531
+ "reasoning_closed": true,
532
+ "seconds": 6.3
533
+ },
534
+ {
535
+ "case": "intervals",
536
+ "effort": "high",
537
+ "seed": 1,
538
+ "clear_thinking": true,
539
+ "note": "",
540
+ "completion_tokens": 676,
541
+ "prompt_tokens": 26,
542
+ "reasoning_chars": 55,
543
+ "content_chars": 2646,
544
+ "finish": "stop",
545
+ "reasoning_closed": true,
546
+ "seconds": 25.2
547
+ },
548
+ {
549
+ "case": "intervals",
550
+ "effort": "max",
551
+ "seed": 1,
552
+ "clear_thinking": true,
553
+ "note": "",
554
+ "completion_tokens": 4292,
555
+ "prompt_tokens": 26,
556
+ "reasoning_chars": 13169,
557
+ "content_chars": 2761,
558
+ "finish": "stop",
559
+ "reasoning_closed": true,
560
+ "seconds": 164.8
561
+ },
562
+ {
563
+ "case": "ssh",
564
+ "effort": "high",
565
+ "seed": 1,
566
+ "clear_thinking": true,
567
+ "note": "",
568
+ "completion_tokens": 839,
569
+ "prompt_tokens": 31,
570
+ "reasoning_chars": 531,
571
+ "content_chars": 2736,
572
+ "finish": "stop",
573
+ "reasoning_closed": true,
574
+ "seconds": 31.5
575
+ },
576
+ {
577
+ "case": "ssh",
578
+ "effort": "max",
579
+ "seed": 1,
580
+ "clear_thinking": true,
581
+ "note": "",
582
+ "completion_tokens": 8317,
583
+ "prompt_tokens": 31,
584
+ "reasoning_chars": 28084,
585
+ "content_chars": 3752,
586
+ "finish": "stop",
587
+ "reasoning_closed": true,
588
+ "seconds": 323.9
589
+ },
590
+ {
591
+ "case": "ratelimiter",
592
+ "effort": "high",
593
+ "seed": 1,
594
+ "clear_thinking": true,
595
+ "note": "",
596
+ "completion_tokens": 5046,
597
+ "prompt_tokens": 47,
598
+ "reasoning_chars": 12146,
599
+ "content_chars": 9032,
600
+ "finish": "stop",
601
+ "reasoning_closed": true,
602
+ "seconds": 195.0
603
+ },
604
+ {
605
+ "case": "ratelimiter",
606
+ "effort": "max",
607
+ "seed": 1,
608
+ "clear_thinking": true,
609
+ "note": "",
610
+ "completion_tokens": 19665,
611
+ "prompt_tokens": 47,
612
+ "reasoning_chars": 62723,
613
+ "content_chars": 14453,
614
+ "finish": "stop",
615
+ "reasoning_closed": true,
616
+ "seconds": 788.2
617
+ },
618
+ {
619
+ "case": "ratelimiter",
620
+ "effort": "low",
621
+ "seed": 1,
622
+ "clear_thinking": true,
623
+ "note": "",
624
+ "completion_tokens": 2004,
625
+ "prompt_tokens": 47,
626
+ "reasoning_chars": 288,
627
+ "content_chars": 7813,
628
+ "finish": "stop",
629
+ "reasoning_closed": true,
630
+ "seconds": 78.0
631
+ },
632
+ {
633
+ "case": "isp",
634
+ "effort": "high",
635
+ "seed": 1,
636
+ "clear_thinking": true,
637
+ "note": "",
638
+ "completion_tokens": 24000,
639
+ "prompt_tokens": 70,
640
+ "reasoning_chars": 74034,
641
+ "content_chars": 10537,
642
+ "finish": "length",
643
+ "reasoning_closed": true,
644
+ "seconds": 982.3
645
+ },
646
+ {
647
+ "case": "isp",
648
+ "effort": "high",
649
+ "seed": 2,
650
+ "clear_thinking": true,
651
+ "note": "",
652
+ "completion_tokens": 24000,
653
+ "prompt_tokens": 70,
654
+ "reasoning_chars": 58939,
655
+ "content_chars": 25611,
656
+ "finish": "length",
657
+ "reasoning_closed": true,
658
+ "seconds": 987.6
659
+ },
660
+ {
661
+ "case": "multiturn",
662
+ "effort": "high",
663
+ "seed": 11,
664
+ "clear_thinking": true,
665
+ "note": "turn 1 OK",
666
+ "completion_tokens": 93,
667
+ "prompt_tokens": 44,
668
+ "reasoning_chars": 59,
669
+ "content_chars": 122,
670
+ "finish": "stop",
671
+ "reasoning_closed": true,
672
+ "seconds": 4.3
673
+ },
674
+ {
675
+ "case": "multiturn",
676
+ "effort": "high",
677
+ "seed": 12,
678
+ "clear_thinking": true,
679
+ "note": "turn 2 OK",
680
+ "completion_tokens": 79,
681
+ "prompt_tokens": 105,
682
+ "reasoning_chars": 47,
683
+ "content_chars": 158,
684
+ "finish": "stop",
685
+ "reasoning_closed": true,
686
+ "seconds": 3.8
687
+ },
688
+ {
689
+ "case": "multiturn",
690
+ "effort": "high",
691
+ "seed": 13,
692
+ "clear_thinking": true,
693
+ "note": "turn 3 OK",
694
+ "completion_tokens": 258,
695
+ "prompt_tokens": 178,
696
+ "reasoning_chars": 141,
697
+ "content_chars": 483,
698
+ "finish": "stop",
699
+ "reasoning_closed": true,
700
+ "seconds": 10.7
701
+ },
702
+ {
703
+ "case": "multiturn",
704
+ "effort": "high",
705
+ "seed": 14,
706
+ "clear_thinking": true,
707
+ "note": "turn 4 OK",
708
+ "completion_tokens": 149,
709
+ "prompt_tokens": 376,
710
+ "reasoning_chars": 77,
711
+ "content_chars": 417,
712
+ "finish": "stop",
713
+ "reasoning_closed": true,
714
+ "seconds": 6.9
715
+ },
716
+ {
717
+ "case": "multiturn",
718
+ "effort": "high",
719
+ "seed": 11,
720
+ "clear_thinking": false,
721
+ "note": "turn 1 OK",
722
+ "completion_tokens": 93,
723
+ "prompt_tokens": 44,
724
+ "reasoning_chars": 59,
725
+ "content_chars": 122,
726
+ "finish": "stop",
727
+ "reasoning_closed": true,
728
+ "seconds": 3.7
729
+ },
730
+ {
731
+ "case": "multiturn",
732
+ "effort": "high",
733
+ "seed": 12,
734
+ "clear_thinking": false,
735
+ "note": "turn 2 OK",
736
+ "completion_tokens": 72,
737
+ "prompt_tokens": 143,
738
+ "reasoning_chars": 45,
739
+ "content_chars": 139,
740
+ "finish": "stop",
741
+ "reasoning_closed": true,
742
+ "seconds": 3.5
743
+ },
744
+ {
745
+ "case": "multiturn",
746
+ "effort": "high",
747
+ "seed": 13,
748
+ "clear_thinking": false,
749
+ "note": "turn 3 OK",
750
+ "completion_tokens": 232,
751
+ "prompt_tokens": 232,
752
+ "reasoning_chars": 186,
753
+ "content_chars": 392,
754
+ "finish": "stop",
755
+ "reasoning_closed": true,
756
+ "seconds": 9.8
757
+ },
758
+ {
759
+ "case": "multiturn",
760
+ "effort": "high",
761
+ "seed": 14,
762
+ "clear_thinking": false,
763
+ "note": "turn 4 OK",
764
+ "completion_tokens": 147,
765
+ "prompt_tokens": 484,
766
+ "reasoning_chars": 91,
767
+ "content_chars": 317,
768
+ "finish": "stop",
769
+ "reasoning_closed": true,
770
+ "seconds": 6.9
771
+ },
772
+ {
773
+ "case": "isp",
774
+ "effort": "max",
775
+ "seed": 1,
776
+ "clear_thinking": true,
777
+ "note": "",
778
+ "completion_tokens": 40000,
779
+ "prompt_tokens": 70,
780
+ "reasoning_chars": 110194,
781
+ "content_chars": 28150,
782
+ "finish": "length",
783
+ "reasoning_closed": true,
784
+ "seconds": 1647.7
785
+ }
786
+ ]
jang_config.json CHANGED
@@ -230,5 +230,17 @@
230
  "tool_parser": "glm_xml_args",
231
  "tool_response_role": "observation",
232
  "family": "glm5_next"
 
 
 
 
 
 
 
 
 
 
 
 
233
  }
234
  }
 
230
  "tool_parser": "glm_xml_args",
231
  "tool_response_role": "observation",
232
  "family": "glm5_next"
233
+ },
234
+ "scale_correction": {
235
+ "revision": 2,
236
+ "rule": "row scale *= (<w,w>/<w,q>)^pow, clipped to [0.5, 2.0]",
237
+ "pow": 1.0,
238
+ "projections": [
239
+ "gate",
240
+ "up",
241
+ "down"
242
+ ],
243
+ "reason": "MSE-optimal scales attenuate deep features; close-reasoning decision restored",
244
+ "codes_changed": false
245
  }
246
  }
model-00003-of-00019.safetensors CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:98e0184dab6b90c7aea114bd61c66ff3f5258df1c2702f83bc1c1a30fac98ad1
3
  size 5446436616
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:4a443e0b29dc02c1ac57edc1254c1a66fa7c0e468259fb8d69ec17cdb4cc31a1
3
  size 5446436616
model-00004-of-00019.safetensors CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:75c7f9790c64b6b749301da42ddc06ddff40a69e448c9380dbacb25b464ba758
3
  size 5748426504
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:7cbb3cd42b4f5cbb4beae4f5dc3ca0b225225ff613f988b36e4b03da36a07f43
3
  size 5748426504
model-00005-of-00019.safetensors CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:67e1d3b2f7ed95eb5c5f6b11cf4e021c0e1dc3e23ba9389b42cf57fd6f7d2725
3
  size 5448796168
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f038e22c7c43f33776ff5ccd20ec871913b84b25c43a41ba5673b94088c0b144
3
  size 5448796168
model-00006-of-00019.safetensors CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:f48888f0a57c25df9f5b9e4e49976632a0dc39098307c11c9348b09bc17caaff
3
  size 5750786072
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:fddda631a6bde61b7bab877f36a4db100c2f3706d985bd614de60df9538e54e7
3
  size 5750786072
model-00007-of-00019.safetensors CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:5a2b119b1a92e519d2d746f0dfdb7cde1c6a3c47cdf51077efad4cc319cb310f
3
  size 5749606424
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:105c31daa7e805fa8417a201a1d1d7053ceb4cbbc9d0a971485ef5f8ec9e215e
3
  size 5749606424
model-00008-of-00019.safetensors CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:ca382f6567c55a7847a73f3ed38bc0460b7e2343bbb18a7ef7b41743b840d006
3
  size 5448796184
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c0c0f0265185ac520771d551d04c088282c2764d2389ca2e8ffcbf54f55ab2a2
3
  size 5448796184
model-00009-of-00019.safetensors CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:165cd20187ecb8bba5a2f46e4dbf0fb959fdc74ec4f624ce3703620254f337be
3
  size 5750786072
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b25b35d7c055014ef8a562e417578152890c4b916a6a75ad04f838b3ac805a90
3
  size 5750786072
model-00010-of-00019.safetensors CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:fa42f244051ddf2331f51f9a037664668fb5cd7ed0870de374350da921308ee2
3
  size 5749606424
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:96dcc156d6b2f786f0e2852a10b287dc87e35eaa17809ee7666aed57cecb23da
3
  size 5749606424
model-00011-of-00019.safetensors CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:9d9b412e7a0d3e6bc13a1ce443b500bd5a82430259ca9bf5384d2f84f2a0a7c4
3
  size 6053955864
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:fd10e26e97cfdaf52d4e48d7ae6c4da4fba952b2f5a1386d57389d39242b68d8
3
  size 6053955864
model-00012-of-00019.safetensors CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:ae2c0ff6613eafdb953c1451947e82b0a8270a313d5dba79e11a4026370a0c3e
3
  size 5448796184
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:7c873f99ea638bd91e963d462e37dc342d092dcbf0d798837ca9dd2a336d0061
3
  size 5448796184
model-00013-of-00019.safetensors CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:e09d144705e55560ba0af5a2e500cc9ab3579d52ad7976c585355da1cee93a9b
3
  size 6052775960
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e11c294709629d0b2b43bdff10490a0c23098d5c37ba68e2ed5060c01037478b
3
  size 6052775960
model-00014-of-00019.safetensors CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:1521a3910f0235f9149e541b4857f0d2ee652cc36a38c7fc0d3a2798ff8829dd
3
  size 5748426512
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:9085b81fd123da17b7171e25fbedb65573d43d0263ab0ae94fbbbfe8361cb895
3
  size 5748426512
model-00015-of-00019.safetensors CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:52fd5bc0b3df7710d07de45f5430f740e0b8d2926ed85a686744a092ce5ec14b
3
  size 5445256720
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:bf111a4f575d3bcc3646b74816ad3e31992a4a918c7721be848968d32370cca0
3
  size 5445256720
model-00016-of-00019.safetensors CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:651b8eb9af1813b4aa079873c6a2af055ed2a98235ffdce767604ecb946eb16e
3
  size 5750786072
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a748a608c1cff05de4b838f2df68b91c629c9faa737b43a1a62fb8f596072559
3
  size 5750786072
model-00017-of-00019.safetensors CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:4fac45d23cae0c60077642480f63ea32e0d42be83f7af1293c235fc6bea9a6e9
3
  size 5749606424
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:2d076f82c8bd7ad3ce2fe338138df5b423ba771d6bf35e27f1e9c120f8953c62
3
  size 5749606424
model-00018-of-00019.safetensors CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:369afe6238c53d206cc87b7b022c944d295e07503cf80e1294e5c92405877a27
3
  size 5448796184
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:0f0116dc141186fe75282e89e229364464f7caa861bb9dc7327154d73ef01b13
3
  size 5448796184
model-00019-of-00019.safetensors CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:fa977a7cc241d7c175a5f2d1f3cb11f2736f2671e61cb412aeeb7bf33084536d
3
  size 909508992
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:44e54e0558dfc97740e3e3baad0df0276bdde2b0bbd35fee30ca3bda0740febd
3
  size 909508992