Hugging Face's logo

jaredpalmer
/
kev-0.8b

Text Classification
PEFT
Safetensors
English
decision-model
calibration
lora
multiple-choice
typesafe
qwen3.5
Eval Results (legacy)
Model card Files
xet
 Community
1
jaredpalmer commited on
Commit
2256796
·
verified ·
1 Parent(s): c917ede

Kev-0.8B: dates+unknowable delta (locked test 0.834 / 0.684); previous weights at tag v7-base

Browse files
Files changed (8) hide show
  1. README.md +21 -20
  2. adapter_model.safetensors +1 -1
  3. head.pt +2 -2
  4. provenance.json +19 -16
  5. result.json +0 -0
  6. train.log +47 -317
  7. training_config.json +14 -4
  8. training_metrics.json +7 -7
README.md CHANGED
@@ -33,54 +33,55 @@ model-index:
33
  - task: { type: text-classification, name: typed decision (choice / noul / score) }
34
  dataset: { type: mixed, name: "decision-v7 development (1,204 records; ten trained public sources + programmatic policy data)" }
35
  metrics:
36
- - { type: accuracy, value: 0.829 }
37
- - { type: expected_calibration_error, value: 0.095, name: "ECE, raw probabilities" }
38
  - task: { type: text-classification, name: typed decision, out-of-domain }
39
  dataset: { type: mixed, name: "transfer-v4 development (764 records; six never-trained sources + held-out policy structures)" }
40
  metrics:
41
- - { type: accuracy, value: 0.643 }
42
- - { type: brier_score, value: 0.513 }
43
  ---
44
 
45
  # Kev-0.8B
46
 
47
  Kev-0.8B is a **decision model**: one document (the *state*) and a set of typed questions in, a probability distribution per question out, in one forward pass. No text generation. It is a LoRA adapter (r=16, 11.3M trainable parameters) plus a pointer head on `Qwen/Qwen3.5-0.8B-Base` (revision `dc7cdfe2`), serving TypeSafe's public `/v1/systemone` contract.
48
 
49
- **The small member of the Kev family.** Same data and recipe as the 0.6B it replaces, on the Qwen3.5 base: in-distribution 0.829 (Kev-0.6B 0.801), out of domain 0.643 (0.620), and it is the first small Kev that learns any rule composition (held-out pairs 0.38 vs 0.08). Three seeds: transfer 0.622 / 0.634 / **0.643**; this checkpoint is seed 2, selected on the development partition (highest development accuracy). Out of domain it is still a sub-1B model: use Kev-4B for accuracy; use this one where memory rules the 4B out, and measure on your own data.
50
 
51
- - Hub: `jaredpalmer/kev-0.8b` (this repo; trial `q35-08b/02-trial-2`)
52
  - Code, suites, results, and the full research log: [github.com/jaredpalmer/kev](https://github.com/jaredpalmer/kev) — `PLAN_Qwen35.md`, `PLAN.md`, `runs/leaderboard.md`
53
 
54
  ## Results (same frozen items for every row)
55
 
56
  | | Kev-0.6B (Qwen3) | **Kev-0.8B** | Kev-4B | Kev-9B | Jev |
57
  |---|---|---|---|---|---|
58
- | in-distribution accuracy (decision-v7 dev, 1,204 records) | 0.801 | **0.829** | 0.877 | 0.876 | 0.845 |
59
- | out-of-domain accuracy (transfer-v4 dev, 764 records) | 0.620 | **0.643** | 0.794 | 0.812 | 0.857 |
60
- | out-of-domain Brier | 0.536 | **0.513** | 0.316 | 0.291 | 0.211 |
61
- | confident errors out of domain (p ≥ 0.9 and wrong) | 10.8% | 9.6% | 8.2% | 7.5% | 3.7% |
62
- | coverage at ≤ 5% error (share of decisions automatable) | – | 0.17 | 0.54 | 0.53 | 0.70 |
63
- | held-out policy structures, both siblings correct | 0.08 | **0.38** | 0.78 | 0.80 | 0.86 |
64
- | option-order flip rate | 0.07 | 0.06 | 0.08 | 0.03 | 0.00 |
65
- | none-option present, accuracy | 0.80 | 0.83 | 0.93 | 0.90 | – |
 
66
 
67
- Per-source out-of-domain accuracy (Kev-0.8B / Jev): QNLI 0.82 / 0.93, SciQ 0.90 / 0.99, TweetEval-offensive 0.62 / 0.81, PAWS 0.59 / 0.79, MMLU 0.41 / 0.90, Emotion 0.54 / 0.59, authorization 0.90 / 1.00, deadline (3-level date arithmetic) 0.28 / 0.93, (A or B) and C 0.66 / 0.91, (A and B) or not C 0.62 / 0.97, if A then not B else C 0.72 / 0.78.
68
 
69
- Paired against Kev-0.6B on the same items (record-clustered bootstrap): +5.7 pp [+1.2, +10.0] out of domain.
70
 
71
- **Locked test, read once** (`runs/locked/kev-08b-q35-ungated/`): in-distribution **0.827** (Brier 0.258, ECE 0.096), out-of-domain **0.668** (Brier 0.473, ECE 0.164, confident errors 9.6%, held-out pairs 0.36). Kev-0.6B on the same test items: 0.808 / 0.642. This partition will not be read again for this checkpoint.
72
 
73
  ## Known limits
74
 
75
  - **Out of domain it is a sub-1B model.** Knowledge (MMLU 0.41) and paraphrase (PAWS 0.59) are near the untrained base; the same recipe reaches 0.79 at 4B and 0.81 at 9B on these items.
76
  - **Slow on a Mac for its size.** The DeltaNet kernels have no MPS implementation; a five-question request takes ~0.33 s in bf16 on an M5 (Kev-0.6B: 0.12 s). On CUDA with `flash-linear-attention` it is fast.
77
  - Requires `transformers >= 5.17` and `peft >= 0.21`.
78
- - Ordinal hedging on date arithmetic (`deadline` 0.28): collapses to the middle level.
79
- - Confident-error rate out of domain is 9.6%; raw ECE 0.095 in-domain, 0.184 out of domain. Probabilities are usable in-domain; treat them as advisory elsewhere.
80
 
81
  ## Training
82
 
83
- Frozen suite `evals/v7/decision-v7`: 10,000 public records (1,000 per source), 896 policy minimal-pair records over nine template families, 1,680 records from 60 randomly generated rule structures in four rendering styles. Two epochs, LoRA r=16 α=32 on attention, MLP and DeltaNet projections; pointer head from scratch; cross-entropy on the option distribution; lr 1e-4 (OneCycle), batch 8, bf16 autocast with fp32 master weights; option permutation, none-of-the-above insertion, distractors, none minimal pairs on 25% of Choice records; ~20 min on one H100. No Jev outputs were used for training.
84
 
85
  ## Evaluation protocol
86
 
 
33
  - task: { type: text-classification, name: typed decision (choice / noul / score) }
34
  dataset: { type: mixed, name: "decision-v7 development (1,204 records; ten trained public sources + programmatic policy data)" }
35
  metrics:
36
+ - { type: accuracy, value: 0.825 }
37
+ - { type: expected_calibration_error, value: 0.110, name: "ECE, raw probabilities" }
38
  - task: { type: text-classification, name: typed decision, out-of-domain }
39
  dataset: { type: mixed, name: "transfer-v4 development (764 records; six never-trained sources + held-out policy structures)" }
40
  metrics:
41
+ - { type: accuracy, value: 0.652 }
42
+ - { type: brier_score, value: 0.499 }
43
  ---
44
 
45
  # Kev-0.8B
46
 
47
  Kev-0.8B is a **decision model**: one document (the *state*) and a set of typed questions in, a probability distribution per question out, in one forward pass. No text generation. It is a LoRA adapter (r=16, 11.3M trainable parameters) plus a pointer head on `Qwen/Qwen3.5-0.8B-Base` (revision `dc7cdfe2`), serving TypeSafe's public `/v1/systemone` contract.
48
 
49
+ **The small member of the Kev family.** Same data and recipe as the 0.6B it replaces, on the Qwen3.5 base: in-distribution 0.825 (Kev-0.6B 0.801), out of domain 0.652 (0.620), and it is the first small Kev that learns any rule composition (held-out pairs 0.42 vs 0.08). Three seeds of the base recipe: transfer 0.622 / 0.634 / **0.643**; this checkpoint is seed 2 (selected on development accuracy) followed by a 9-minute **delta fine-tune** on 1,425 generated records (date-bearing policy cases with explicit day counts; evidence-free cases with uniform targets) mixed with 2,000 replayed training records — the same delta as Kev-4B and Kev-9B. Locked test against the pre-delta checkpoint: out of domain 0.668 → **0.684** (+2.2 pp [−0.8, +5.5]), Brier 0.473 → 0.460. Out of domain it is still a sub-1B model: use Kev-4B for accuracy; use this one where memory rules the 4B out, and measure on your own data.
50
 
51
+ - Hub: `jaredpalmer/kev-0.8b` (this repo; trial `night2-08b-du2/00-trial-0`). The pre-delta checkpoint is at revision `v7-base`.
52
  - Code, suites, results, and the full research log: [github.com/jaredpalmer/kev](https://github.com/jaredpalmer/kev) — `PLAN_Qwen35.md`, `PLAN.md`, `runs/leaderboard.md`
53
 
54
  ## Results (same frozen items for every row)
55
 
56
  | | Kev-0.6B (Qwen3) | **Kev-0.8B** | Kev-4B | Kev-9B | Jev |
57
  |---|---|---|---|---|---|
58
+ | in-distribution accuracy (decision-v7 dev, 1,204 records) | 0.801 | **0.825** | 0.872 | 0.872 | 0.845 |
59
+ | out-of-domain accuracy (transfer-v4 dev, 764 records) | 0.620 | **0.652** | 0.797 | 0.822 | 0.857 |
60
+ | out-of-domain Brier | 0.536 | **0.499** | 0.299 | 0.286 | 0.211 |
61
+ | confident errors out of domain (p ≥ 0.9 and wrong) | 10.8% | 9.9% | 6.9% | 8.7% | 3.7% |
62
+ | coverage at ≤ 5% error (share of decisions automatable) | – | 0.23 | 0.54 | 0.47 | 0.70 |
63
+ | held-out policy structures, both siblings correct | 0.08 | **0.42** | 0.78 | 0.83 | 0.86 |
64
+ | option-order flip rate | 0.07 | 0.08 | 0.08 | 0.03 | 0.00 |
65
+ | none-option present, accuracy | 0.80 | 0.83 | 0.92 | 0.90 | – |
66
+ | calibrated (T = 2.2, fitted in-distribution): Brier / ECE / confident errors | – | 0.433 / 0.062 / 0.8% | | | |
67
 
68
+ Per-source out-of-domain accuracy (Kev-0.8B / Jev): QNLI 0.85 / 0.93, SciQ 0.91 / 0.99, TweetEval-offensive 0.68 / 0.81, PAWS 0.55 / 0.79, MMLU 0.42 / 0.90, Emotion 0.54 / 0.59, authorization 0.97 / 1.00, deadline (3-level date arithmetic) 0.38 / 0.93, (A or B) and C 0.66 / 0.91, (A and B) or not C 0.56 / 0.97, if A then not B else C 0.59 / 0.78.
69
 
70
+ Paired against Kev-0.6B on the same items (record-clustered bootstrap), before the delta: +5.7 pp [+1.2, +10.0] out of domain; the delta adds +0.5 pp [−3.2, +3.8] on development and +2.2 pp on the locked test.
71
 
72
+ **Locked test, read once per checkpoint** (`runs/locked/kev-08b-night2-du-ungated/`; pre-delta `runs/locked/kev-08b-q35-ungated/`): in-distribution **0.834** (Brier 0.268, ECE 0.100), out-of-domain **0.684** (Brier 0.460, ECE 0.154, confident errors 8.7%, held-out pairs 0.45). Pre-delta: 0.827 / 0.668; Kev-0.6B on the same test items: 0.808 / 0.642.
73
 
74
  ## Known limits
75
 
76
  - **Out of domain it is a sub-1B model.** Knowledge (MMLU 0.41) and paraphrase (PAWS 0.59) are near the untrained base; the same recipe reaches 0.79 at 4B and 0.81 at 9B on these items.
77
  - **Slow on a Mac for its size.** The DeltaNet kernels have no MPS implementation; a five-question request takes ~0.33 s in bf16 on an M5 (Kev-0.6B: 0.12 s). On CUDA with `flash-linear-attention` it is fast.
78
  - Requires `transformers >= 5.17` and `peft >= 0.21`.
79
+ - Ordinal hedging on date arithmetic (`deadline` 0.38): collapses to the middle level. `KEV_DATE_FACTS=1` (day counts appended to the state) helps the larger models more than this one.
80
+ - Confident-error rate out of domain is 9.9% raw; `KEV_TEMPERATURE=2.2` brings it to 0.8% and ECE from 0.179 to 0.062 without changing any answer. Probabilities are usable in-domain; treat them as advisory elsewhere.
81
 
82
  ## Training
83
 
84
+ Frozen suite `evals/v7/decision-v7`: 10,000 public records (1,000 per source), 896 policy minimal-pair records over nine template families, 1,680 records from 60 randomly generated rule structures in four rendering styles. Two epochs, LoRA r=16 α=32 on attention, MLP and DeltaNet projections; pointer head from scratch; cross-entropy on the option distribution; lr 1e-4 (OneCycle), batch 8, bf16 autocast with fp32 master weights; option permutation, none-of-the-above insertion, distractors, none minimal pairs on 25% of Choice records; ~20 min on one H100. Then the delta: `--init_from jaredpalmer/kev-0.8b@v7-base --data evals/night2/dates_unknowable.jsonl --replay 2000 --lr 4e-5 --epochs 1`, 9 minutes. No Jev outputs were used for training.
85
 
86
  ## Evaluation protocol
87
 
adapter_model.safetensors CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:d9fa619fd3b0490122454c386bfa1b53c23850189d1b5c4346962162e2e39a64
3
  size 43338624
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c81d5716f0af7622d8d2b97013c333d48263ca01113f9a7cf4526e96f6ac0b26
3
  size 43338624
head.pt CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:8610dac1c30a64bbcea7715f20a258f4d084d7b862c10f96a1625c732049a32a
3
- size 2102399
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:591953b2ae128ef2adcf90124f59ac11fa05dec572525f9a6f0edf883079a995
3
+ size 2102911
provenance.json CHANGED
@@ -1,8 +1,8 @@
1
  {
2
  "config": {
3
- "epochs": 2,
4
- "seed": 2,
5
- "lr": 0.0001,
6
  "lora": 16,
7
  "accum": 1,
8
  "batch": 8,
@@ -18,38 +18,41 @@
18
  "head_lr": 0.0,
19
  "weight_decay": 0.01,
20
  "anchor_w": 0.0,
 
 
21
  "base": "Qwen/Qwen3.5-0.8B-Base",
22
  "base_revision": "dc7cdfe2ee4154fa7e30f5b51ca41bfa40174e68",
23
- "dtype": "bf16"
 
24
  },
25
- "config_sha256": "1f902b8d0af3e9c1384f56f9f248c92d27f7029156a1a640389fb2acf5ea6b0a",
26
  "suite_sha256": "a8f50e481b7d90b97da049e0ff6a01cee2f1ed204aed61a8265af0edbb5514d2",
27
  "source_hashes": {
28
  "kev/__init__.py": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855",
29
  "kev/anchors.py": "089d8a5493502bb26f540eb1c5e681780ca0bb01276073d4e1733211f15d0e10",
30
- "kev/api.py": "cdb0602684d798ccd3fc2c86be622f2a8e7bcdeee64d07f95604c7d3fd4701ec",
31
  "kev/autoresearch.py": "0a8aa6b57c1c9ba93d25b2cf631b03aed2e686e374c1148002eec48765b7b1fc",
32
- "kev/benchmark.py": "5be816b27bc69c7244fd1b0ad619d4783b85a21b928212f306e731312f5146b2",
33
  "kev/compare.py": "bd0445f021de59e35c7bff9304e39dd2e1211e3877a453575594ae7b81b0ada4",
34
  "kev/composition.py": "f335ed17e18e0a544893db5e22b9a059e6ce1b2e14dbb863ac7d7bbf8f3e0536",
35
  "kev/contrastive.py": "cbb979aa5d40265ad0e64695f94b281d91751fa405811ddfa8212fede111edcf",
36
- "kev/data.py": "9729f8f497b02aca542915d6f1bea3ff55956629d7d9610370aea91ee66efd23",
37
- "kev/evaluate.py": "6f95c52f757ecffc887710e1b354f06c69c35336421d3e75eae84d251ddd120c",
38
- "kev/experiment.py": "c635394fe56c9aa11765f8bbf8f133dcf3e57cbe8f2bce4ebdc7b4e85fa536ff",
39
  "kev/jev.py": "e0213782359ba2f95adbf045ddaf0a008b08b4162bf6a9b66d4ce55fc91a51cb",
40
- "kev/model.py": "46d72ec55c9c35d28e8c43d820733e21bc5f3f261ff72a939fe029b88c4fcb1c",
41
  "kev/plot.py": "d689c7dd18cde7f9ea77cb4f55c50ff7da1880b21cb2ecf244348e45842a1382",
42
- "kev/publish.py": "c0b3efac3cfe93990d1846cfd306cf762a9d03e9e58380d13c2961a5ddae4619",
43
- "kev/serve.py": "570e995aa98b51dbbf867694276d29fc9b9f3f1e62f22bea0116ae51d6497003",
44
  "kev/study_v3.py": "9fc44d44dad09a4f1ee29a7bcb2eb3c7aa373d09186d0e9533666b69ec95c401",
45
  "kev/suite.py": "44858f9a99df086a47d1ae36141631fab6fb6a8293a9e1d0c6ad397ddcfaf94a",
46
- "kev/train.py": "e405978b6e1e7444e682a8869d1bd6c88c90bd9890836fcd84b2ab2aa0543817",
47
  "kev/transfer_v9.py": "588aa2ff3ec0823c2e31733bef9a3748b263849c966ef6a80ebd686539c605e9",
48
- "modal_app.py": "12c1b6c1d7f264da8644819bfd11b4baab9c541b1277fee120467514d4628591",
49
  "pyproject.toml": "52da5eea3efc6f2b1c0589acebad62e56a214294bb02c1a4218c93efd4af3182",
50
  "uv.lock": "18b3e5ea0f25d2e8546fab81f16cb965ae05c3289fffaaa1ce27d114adee47f3"
51
  },
52
- "git_commit": "5f78968927069eaacc3b2bdb688586989b3933ac",
53
  "platform": "Linux-4.19.0-gvisor-x86_64-with-glibc2.36",
54
  "torch": "2.8.0+cu128",
55
  "device": "cuda",
 
1
  {
2
  "config": {
3
+ "epochs": 1,
4
+ "seed": 1,
5
+ "lr": 4e-05,
6
  "lora": 16,
7
  "accum": 1,
8
  "batch": 8,
 
18
  "head_lr": 0.0,
19
  "weight_decay": 0.01,
20
  "anchor_w": 0.0,
21
+ "dtype": "bf16",
22
+ "replay": 2000,
23
  "base": "Qwen/Qwen3.5-0.8B-Base",
24
  "base_revision": "dc7cdfe2ee4154fa7e30f5b51ca41bfa40174e68",
25
+ "init_from": "jaredpalmer/kev-0.8b",
26
+ "data": "evals/night2/dates_unknowable.jsonl"
27
  },
28
+ "config_sha256": "0e130e9f338c8a4c911c6dc37493b36e6be7ca972cbafbb42466043b6daa7a4b",
29
  "suite_sha256": "a8f50e481b7d90b97da049e0ff6a01cee2f1ed204aed61a8265af0edbb5514d2",
30
  "source_hashes": {
31
  "kev/__init__.py": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855",
32
  "kev/anchors.py": "089d8a5493502bb26f540eb1c5e681780ca0bb01276073d4e1733211f15d0e10",
33
+ "kev/api.py": "c9eebbdb6c625563d33f030299ce0fbe8b50493dd1e92592df9514c30f5720af",
34
  "kev/autoresearch.py": "0a8aa6b57c1c9ba93d25b2cf631b03aed2e686e374c1148002eec48765b7b1fc",
35
+ "kev/benchmark.py": "4192ec3b26b065452f84bde38a091e6854a28fe185d2e0f39a0c91b2efe69df7",
36
  "kev/compare.py": "bd0445f021de59e35c7bff9304e39dd2e1211e3877a453575594ae7b81b0ada4",
37
  "kev/composition.py": "f335ed17e18e0a544893db5e22b9a059e6ce1b2e14dbb863ac7d7bbf8f3e0536",
38
  "kev/contrastive.py": "cbb979aa5d40265ad0e64695f94b281d91751fa405811ddfa8212fede111edcf",
39
+ "kev/data.py": "997c31d737c2a130ade49edd6534aa47d910786c98af883745b1a97ec858a704",
40
+ "kev/evaluate.py": "1b20e3f9edf417aa8dae924b1526e52f74b710cadf7213c5ec68334f6e7f8fe1",
41
+ "kev/experiment.py": "ba034856f79bd500d8b2eeeb7382e4db3fccb30b437f9f3cf55cf1a7c565a6c9",
42
  "kev/jev.py": "e0213782359ba2f95adbf045ddaf0a008b08b4162bf6a9b66d4ce55fc91a51cb",
43
+ "kev/model.py": "a17d57a52da3fe6ebef7146ce548bd3cedc09e21cc2b7738d9077771f5f16989",
44
  "kev/plot.py": "d689c7dd18cde7f9ea77cb4f55c50ff7da1880b21cb2ecf244348e45842a1382",
45
+ "kev/publish.py": "8f23cda767008756e0f169249eb44577db76bd6c1e12e9cf64c302bad04c8f16",
46
+ "kev/serve.py": "d095bbee3a210fa8807d1a7b073b56181c93aa69b00f2e243a323ab0a19fa8c9",
47
  "kev/study_v3.py": "9fc44d44dad09a4f1ee29a7bcb2eb3c7aa373d09186d0e9533666b69ec95c401",
48
  "kev/suite.py": "44858f9a99df086a47d1ae36141631fab6fb6a8293a9e1d0c6ad397ddcfaf94a",
49
+ "kev/train.py": "c2a3770072b73dd4c0769c8f188ddedbb402f2f8195df6e10b11e5112cab734b",
50
  "kev/transfer_v9.py": "588aa2ff3ec0823c2e31733bef9a3748b263849c966ef6a80ebd686539c605e9",
51
+ "modal_app.py": "d3700b2914be5ef6baa6d588a967d3a242903e1ed96d0f5cc525a8908aa58340",
52
  "pyproject.toml": "52da5eea3efc6f2b1c0589acebad62e56a214294bb02c1a4218c93efd4af3182",
53
  "uv.lock": "18b3e5ea0f25d2e8546fab81f16cb965ae05c3289fffaaa1ce27d114adee47f3"
54
  },
55
+ "git_commit": "19dcae9b6e3e1a48200c5825aad9fc200d31e20a",
56
  "platform": "Linux-4.19.0-gvisor-x86_64-with-glibc2.36",
57
  "torch": "2.8.0+cu128",
58
  "device": "cuda",
result.json CHANGED
The diff for this file is too large to render. See raw diff
 
train.log CHANGED
@@ -1,319 +1,49 @@
 
 
1
  device=cuda trainable params=11.3M
2
- 12576 training requests (holdout=[]), questions by type {'score': 3448, 'noul': 5224, 'choice': 6904}
 
3
  [transformers] `causal_conv1d_fn` is falling back to its reference PyTorch implementation because `causal_conv1d` is not installed. This is correct but much slower; install `causal_conv1d` for the optimized kernel.
4
- ep0 step 10/3144 loss 1.960 kl 0.000 anchor 0.000 1.843s/rec
5
- ep0 step 20/3144 loss 1.867 kl 0.000 anchor 0.000 0.987s/rec
6
- ep0 step 30/3144 loss 1.803 kl 0.000 anchor 0.000 0.660s/rec
7
- ep0 step 40/3144 loss 1.955 kl 0.000 anchor 0.000 0.499s/rec
8
- ep0 step 50/3144 loss 1.721 kl 0.000 anchor 0.000 0.409s/rec
9
- ep0 step 60/3144 loss 1.448 kl 0.000 anchor 0.000 0.348s/rec
10
- ep0 step 70/3144 loss 2.027 kl 0.000 anchor 0.000 0.304s/rec
11
- ep0 step 80/3144 loss 1.461 kl 0.000 anchor 0.000 0.301s/rec
12
- ep0 step 90/3144 loss 1.482 kl 0.000 anchor 0.000 0.268s/rec
13
- ep0 step 100/3144 loss 1.319 kl 0.000 anchor 0.000 0.245s/rec
14
- ep0 step 110/3144 loss 1.504 kl 0.000 anchor 0.000 0.224s/rec
15
- ep0 step 120/3144 loss 1.287 kl 0.000 anchor 0.000 0.210s/rec
16
- ep0 step 130/3144 loss 1.264 kl 0.000 anchor 0.000 0.197s/rec
17
- ep0 step 140/3144 loss 1.123 kl 0.000 anchor 0.000 0.184s/rec
18
- ep0 step 150/3144 loss 1.177 kl 0.000 anchor 0.000 0.173s/rec
19
- ep0 step 160/3144 loss 1.214 kl 0.000 anchor 0.000 0.166s/rec
20
- ep0 step 170/3144 loss 1.112 kl 0.000 anchor 0.000 0.158s/rec
21
- ep0 step 180/3144 loss 0.867 kl 0.000 anchor 0.000 0.153s/rec
22
- ep0 step 190/3144 loss 1.256 kl 0.000 anchor 0.000 0.148s/rec
23
- ep0 step 200/3144 loss 1.032 kl 0.000 anchor 0.000 0.142s/rec
24
- ep0 step 210/3144 loss 1.038 kl 0.000 anchor 0.000 0.137s/rec
25
- ep0 step 220/3144 loss 1.062 kl 0.000 anchor 0.000 0.133s/rec
26
- ep0 step 230/3144 loss 0.815 kl 0.000 anchor 0.000 0.128s/rec
27
- ep0 step 240/3144 loss 0.844 kl 0.000 anchor 0.000 0.124s/rec
28
- ep0 step 250/3144 loss 0.886 kl 0.000 anchor 0.000 0.121s/rec
29
- ep0 step 260/3144 loss 0.806 kl 0.000 anchor 0.000 0.117s/rec
30
- ep0 step 270/3144 loss 0.755 kl 0.000 anchor 0.000 0.114s/rec
31
- ep0 step 280/3144 loss 0.961 kl 0.000 anchor 0.000 0.118s/rec
32
- ep0 step 290/3144 loss 1.129 kl 0.000 anchor 0.000 0.115s/rec
33
- ep0 step 300/3144 loss 0.962 kl 0.000 anchor 0.000 0.113s/rec
34
- ep0 step 310/3144 loss 0.899 kl 0.000 anchor 0.000 0.110s/rec
35
- ep0 step 320/3144 loss 0.744 kl 0.000 anchor 0.000 0.108s/rec
36
- ep0 step 330/3144 loss 0.824 kl 0.000 anchor 0.000 0.105s/rec
37
- ep0 step 340/3144 loss 0.752 kl 0.000 anchor 0.000 0.103s/rec
38
- ep0 step 350/3144 loss 0.817 kl 0.000 anchor 0.000 0.101s/rec
39
- ep0 step 360/3144 loss 0.547 kl 0.000 anchor 0.000 0.099s/rec
40
- ep0 step 370/3144 loss 0.776 kl 0.000 anchor 0.000 0.097s/rec
41
- ep0 step 380/3144 loss 0.734 kl 0.000 anchor 0.000 0.095s/rec
42
- ep0 step 390/3144 loss 0.920 kl 0.000 anchor 0.000 0.094s/rec
43
- ep0 step 400/3144 loss 1.078 kl 0.000 anchor 0.000 0.092s/rec
44
- ep0 step 410/3144 loss 0.803 kl 0.000 anchor 0.000 0.091s/rec
45
- ep0 step 420/3144 loss 0.653 kl 0.000 anchor 0.000 0.089s/rec
46
- ep0 step 430/3144 loss 0.845 kl 0.000 anchor 0.000 0.088s/rec
47
- ep0 step 440/3144 loss 0.781 kl 0.000 anchor 0.000 0.087s/rec
48
- ep0 step 450/3144 loss 0.757 kl 0.000 anchor 0.000 0.085s/rec
49
- ep0 step 460/3144 loss 0.768 kl 0.000 anchor 0.000 0.084s/rec
50
- ep0 step 470/3144 loss 0.727 kl 0.000 anchor 0.000 0.083s/rec
51
- ep0 step 480/3144 loss 0.734 kl 0.000 anchor 0.000 0.082s/rec
52
- ep0 step 490/3144 loss 0.642 kl 0.000 anchor 0.000 0.081s/rec
53
- ep0 step 500/3144 loss 0.758 kl 0.000 anchor 0.000 0.080s/rec
54
- ep0 step 510/3144 loss 0.840 kl 0.000 anchor 0.000 0.079s/rec
55
- ep0 step 520/3144 loss 0.651 kl 0.000 anchor 0.000 0.078s/rec
56
- ep0 step 530/3144 loss 1.135 kl 0.000 anchor 0.000 0.077s/rec
57
- ep0 step 540/3144 loss 0.673 kl 0.000 anchor 0.000 0.076s/rec
58
- ep0 step 550/3144 loss 0.567 kl 0.000 anchor 0.000 0.075s/rec
59
- ep0 step 560/3144 loss 1.174 kl 0.000 anchor 0.000 0.075s/rec
60
- ep0 step 570/3144 loss 0.739 kl 0.000 anchor 0.000 0.074s/rec
61
- ep0 step 580/3144 loss 0.647 kl 0.000 anchor 0.000 0.073s/rec
62
- ep0 step 590/3144 loss 0.707 kl 0.000 anchor 0.000 0.073s/rec
63
- ep0 step 600/3144 loss 0.629 kl 0.000 anchor 0.000 0.072s/rec
64
- ep0 step 610/3144 loss 0.530 kl 0.000 anchor 0.000 0.072s/rec
65
- ep0 step 620/3144 loss 0.579 kl 0.000 anchor 0.000 0.071s/rec
66
- ep0 step 630/3144 loss 0.710 kl 0.000 anchor 0.000 0.071s/rec
67
- ep0 step 640/3144 loss 0.795 kl 0.000 anchor 0.000 0.070s/rec
68
- ep0 step 650/3144 loss 0.729 kl 0.000 anchor 0.000 0.070s/rec
69
- ep0 step 660/3144 loss 0.806 kl 0.000 anchor 0.000 0.069s/rec
70
- ep0 step 670/3144 loss 0.847 kl 0.000 anchor 0.000 0.068s/rec
71
- ep0 step 680/3144 loss 0.735 kl 0.000 anchor 0.000 0.068s/rec
72
- ep0 step 690/3144 loss 0.663 kl 0.000 anchor 0.000 0.067s/rec
73
- ep0 step 700/3144 loss 0.755 kl 0.000 anchor 0.000 0.067s/rec
74
- ep0 step 710/3144 loss 0.572 kl 0.000 anchor 0.000 0.066s/rec
75
- ep0 step 720/3144 loss 0.413 kl 0.000 anchor 0.000 0.066s/rec
76
- ep0 step 730/3144 loss 0.668 kl 0.000 anchor 0.000 0.065s/rec
77
- ep0 step 740/3144 loss 0.540 kl 0.000 anchor 0.000 0.065s/rec
78
- ep0 step 750/3144 loss 0.683 kl 0.000 anchor 0.000 0.064s/rec
79
- ep0 step 760/3144 loss 0.517 kl 0.000 anchor 0.000 0.064s/rec
80
- ep0 step 770/3144 loss 0.762 kl 0.000 anchor 0.000 0.064s/rec
81
- ep0 step 780/3144 loss 0.634 kl 0.000 anchor 0.000 0.063s/rec
82
- ep0 step 790/3144 loss 0.612 kl 0.000 anchor 0.000 0.063s/rec
83
- ep0 step 800/3144 loss 0.746 kl 0.000 anchor 0.000 0.062s/rec
84
- ep0 step 810/3144 loss 0.471 kl 0.000 anchor 0.000 0.062s/rec
85
- ep0 step 820/3144 loss 0.620 kl 0.000 anchor 0.000 0.061s/rec
86
- ep0 step 830/3144 loss 0.470 kl 0.000 anchor 0.000 0.061s/rec
87
- ep0 step 840/3144 loss 0.836 kl 0.000 anchor 0.000 0.061s/rec
88
- ep0 step 850/3144 loss 0.670 kl 0.000 anchor 0.000 0.060s/rec
89
- ep0 step 860/3144 loss 0.489 kl 0.000 anchor 0.000 0.060s/rec
90
- ep0 step 870/3144 loss 0.845 kl 0.000 anchor 0.000 0.060s/rec
91
- ep0 step 880/3144 loss 0.760 kl 0.000 anchor 0.000 0.059s/rec
92
- ep0 step 890/3144 loss 0.577 kl 0.000 anchor 0.000 0.059s/rec
93
- ep0 step 900/3144 loss 0.841 kl 0.000 anchor 0.000 0.059s/rec
94
- ep0 step 910/3144 loss 0.743 kl 0.000 anchor 0.000 0.058s/rec
95
- ep0 step 920/3144 loss 0.477 kl 0.000 anchor 0.000 0.058s/rec
96
- ep0 step 930/3144 loss 0.438 kl 0.000 anchor 0.000 0.058s/rec
97
- ep0 step 940/3144 loss 1.025 kl 0.000 anchor 0.000 0.058s/rec
98
- ep0 step 950/3144 loss 0.442 kl 0.000 anchor 0.000 0.057s/rec
99
- ep0 step 960/3144 loss 1.030 kl 0.000 anchor 0.000 0.057s/rec
100
- ep0 step 970/3144 loss 0.694 kl 0.000 anchor 0.000 0.057s/rec
101
- ep0 step 980/3144 loss 0.540 kl 0.000 anchor 0.000 0.057s/rec
102
- ep0 step 990/3144 loss 0.481 kl 0.000 anchor 0.000 0.056s/rec
103
- ep0 step 1000/3144 loss 0.904 kl 0.000 anchor 0.000 0.056s/rec
104
- ep0 step 1010/3144 loss 0.575 kl 0.000 anchor 0.000 0.056s/rec
105
- ep0 step 1020/3144 loss 0.900 kl 0.000 anchor 0.000 0.056s/rec
106
- ep0 step 1030/3144 loss 0.500 kl 0.000 anchor 0.000 0.055s/rec
107
- ep0 step 1040/3144 loss 0.616 kl 0.000 anchor 0.000 0.055s/rec
108
- ep0 step 1050/3144 loss 0.475 kl 0.000 anchor 0.000 0.055s/rec
109
- ep0 step 1060/3144 loss 0.551 kl 0.000 anchor 0.000 0.055s/rec
110
- ep0 step 1070/3144 loss 0.524 kl 0.000 anchor 0.000 0.055s/rec
111
- ep0 step 1080/3144 loss 0.647 kl 0.000 anchor 0.000 0.054s/rec
112
- ep0 step 1090/3144 loss 0.662 kl 0.000 anchor 0.000 0.054s/rec
113
- ep0 step 1100/3144 loss 0.659 kl 0.000 anchor 0.000 0.054s/rec
114
- ep0 step 1110/3144 loss 0.671 kl 0.000 anchor 0.000 0.054s/rec
115
- ep0 step 1120/3144 loss 0.542 kl 0.000 anchor 0.000 0.053s/rec
116
- ep0 step 1130/3144 loss 0.498 kl 0.000 anchor 0.000 0.053s/rec
117
- ep0 step 1140/3144 loss 0.415 kl 0.000 anchor 0.000 0.053s/rec
118
- ep0 step 1150/3144 loss 0.540 kl 0.000 anchor 0.000 0.053s/rec
119
- ep0 step 1160/3144 loss 0.790 kl 0.000 anchor 0.000 0.053s/rec
120
- ep0 step 1170/3144 loss 0.597 kl 0.000 anchor 0.000 0.052s/rec
121
- ep0 step 1180/3144 loss 0.521 kl 0.000 anchor 0.000 0.052s/rec
122
- ep0 step 1190/3144 loss 0.559 kl 0.000 anchor 0.000 0.052s/rec
123
- ep0 step 1200/3144 loss 0.601 kl 0.000 anchor 0.000 0.052s/rec
124
- ep0 step 1210/3144 loss 0.420 kl 0.000 anchor 0.000 0.052s/rec
125
- ep0 step 1220/3144 loss 0.577 kl 0.000 anchor 0.000 0.052s/rec
126
- ep0 step 1230/3144 loss 0.510 kl 0.000 anchor 0.000 0.051s/rec
127
- ep0 step 1240/3144 loss 0.596 kl 0.000 anchor 0.000 0.051s/rec
128
- ep0 step 1250/3144 loss 0.649 kl 0.000 anchor 0.000 0.051s/rec
129
- ep0 step 1260/3144 loss 0.590 kl 0.000 anchor 0.000 0.051s/rec
130
- ep0 step 1270/3144 loss 0.610 kl 0.000 anchor 0.000 0.051s/rec
131
- ep0 step 1280/3144 loss 0.623 kl 0.000 anchor 0.000 0.051s/rec
132
- ep0 step 1290/3144 loss 0.706 kl 0.000 anchor 0.000 0.051s/rec
133
- ep0 step 1300/3144 loss 0.876 kl 0.000 anchor 0.000 0.050s/rec
134
- ep0 step 1310/3144 loss 0.552 kl 0.000 anchor 0.000 0.050s/rec
135
- ep0 step 1320/3144 loss 0.539 kl 0.000 anchor 0.000 0.050s/rec
136
- ep0 step 1330/3144 loss 0.888 kl 0.000 anchor 0.000 0.050s/rec
137
- ep0 step 1340/3144 loss 0.550 kl 0.000 anchor 0.000 0.050s/rec
138
- ep0 step 1350/3144 loss 0.407 kl 0.000 anchor 0.000 0.050s/rec
139
- ep0 step 1360/3144 loss 0.783 kl 0.000 anchor 0.000 0.050s/rec
140
- ep0 step 1370/3144 loss 0.627 kl 0.000 anchor 0.000 0.049s/rec
141
- ep0 step 1380/3144 loss 0.720 kl 0.000 anchor 0.000 0.049s/rec
142
- ep0 step 1390/3144 loss 0.507 kl 0.000 anchor 0.000 0.049s/rec
143
- ep0 step 1400/3144 loss 0.719 kl 0.000 anchor 0.000 0.049s/rec
144
- ep0 step 1410/3144 loss 0.376 kl 0.000 anchor 0.000 0.049s/rec
145
- ep0 step 1420/3144 loss 0.633 kl 0.000 anchor 0.000 0.049s/rec
146
- ep0 step 1430/3144 loss 0.462 kl 0.000 anchor 0.000 0.049s/rec
147
- ep0 step 1440/3144 loss 0.511 kl 0.000 anchor 0.000 0.048s/rec
148
- ep0 step 1450/3144 loss 0.807 kl 0.000 anchor 0.000 0.048s/rec
149
- ep0 step 1460/3144 loss 0.682 kl 0.000 anchor 0.000 0.048s/rec
150
- ep0 step 1470/3144 loss 0.530 kl 0.000 anchor 0.000 0.048s/rec
151
- ep0 step 1480/3144 loss 0.480 kl 0.000 anchor 0.000 0.048s/rec
152
- ep0 step 1490/3144 loss 0.479 kl 0.000 anchor 0.000 0.048s/rec
153
- ep0 step 1500/3144 loss 0.635 kl 0.000 anchor 0.000 0.048s/rec
154
- ep0 step 1510/3144 loss 0.489 kl 0.000 anchor 0.000 0.048s/rec
155
- ep0 step 1520/3144 loss 0.695 kl 0.000 anchor 0.000 0.048s/rec
156
- ep0 step 1530/3144 loss 0.482 kl 0.000 anchor 0.000 0.047s/rec
157
- ep0 step 1540/3144 loss 0.534 kl 0.000 anchor 0.000 0.047s/rec
158
- ep0 step 1550/3144 loss 0.459 kl 0.000 anchor 0.000 0.047s/rec
159
- ep0 step 1560/3144 loss 0.750 kl 0.000 anchor 0.000 0.047s/rec
160
- ep0 step 1570/3144 loss 0.667 kl 0.000 anchor 0.000 0.047s/rec
161
- ep1 step 1580/3144 loss 0.536 kl 0.000 anchor 0.000 0.047s/rec
162
- ep1 step 1590/3144 loss 0.318 kl 0.000 anchor 0.000 0.047s/rec
163
- ep1 step 1600/3144 loss 0.469 kl 0.000 anchor 0.000 0.047s/rec
164
- ep1 step 1610/3144 loss 0.257 kl 0.000 anchor 0.000 0.047s/rec
165
- ep1 step 1620/3144 loss 0.456 kl 0.000 anchor 0.000 0.046s/rec
166
- ep1 step 1630/3144 loss 0.237 kl 0.000 anchor 0.000 0.046s/rec
167
- ep1 step 1640/3144 loss 0.305 kl 0.000 anchor 0.000 0.046s/rec
168
- ep1 step 1650/3144 loss 0.581 kl 0.000 anchor 0.000 0.046s/rec
169
- ep1 step 1660/3144 loss 0.370 kl 0.000 anchor 0.000 0.046s/rec
170
- ep1 step 1670/3144 loss 0.515 kl 0.000 anchor 0.000 0.046s/rec
171
- ep1 step 1680/3144 loss 0.491 kl 0.000 anchor 0.000 0.046s/rec
172
- ep1 step 1690/3144 loss 0.393 kl 0.000 anchor 0.000 0.046s/rec
173
- ep1 step 1700/3144 loss 0.534 kl 0.000 anchor 0.000 0.046s/rec
174
- ep1 step 1710/3144 loss 0.379 kl 0.000 anchor 0.000 0.046s/rec
175
- ep1 step 1720/3144 loss 0.960 kl 0.000 anchor 0.000 0.046s/rec
176
- ep1 step 1730/3144 loss 0.450 kl 0.000 anchor 0.000 0.046s/rec
177
- ep1 step 1740/3144 loss 0.288 kl 0.000 anchor 0.000 0.045s/rec
178
- ep1 step 1750/3144 loss 0.359 kl 0.000 anchor 0.000 0.045s/rec
179
- ep1 step 1760/3144 loss 0.368 kl 0.000 anchor 0.000 0.045s/rec
180
- ep1 step 1770/3144 loss 0.423 kl 0.000 anchor 0.000 0.045s/rec
181
- ep1 step 1780/3144 loss 0.353 kl 0.000 anchor 0.000 0.045s/rec
182
- ep1 step 1790/3144 loss 0.357 kl 0.000 anchor 0.000 0.045s/rec
183
- ep1 step 1800/3144 loss 0.367 kl 0.000 anchor 0.000 0.045s/rec
184
- ep1 step 1810/3144 loss 0.412 kl 0.000 anchor 0.000 0.045s/rec
185
- ep1 step 1820/3144 loss 0.309 kl 0.000 anchor 0.000 0.045s/rec
186
- ep1 step 1830/3144 loss 0.460 kl 0.000 anchor 0.000 0.045s/rec
187
- ep1 step 1840/3144 loss 0.411 kl 0.000 anchor 0.000 0.045s/rec
188
- ep1 step 1850/3144 loss 0.449 kl 0.000 anchor 0.000 0.045s/rec
189
- ep1 step 1860/3144 loss 0.650 kl 0.000 anchor 0.000 0.044s/rec
190
- ep1 step 1870/3144 loss 0.326 kl 0.000 anchor 0.000 0.044s/rec
191
- ep1 step 1880/3144 loss 0.416 kl 0.000 anchor 0.000 0.044s/rec
192
- ep1 step 1890/3144 loss 0.279 kl 0.000 anchor 0.000 0.044s/rec
193
- ep1 step 1900/3144 loss 0.542 kl 0.000 anchor 0.000 0.044s/rec
194
- ep1 step 1910/3144 loss 0.401 kl 0.000 anchor 0.000 0.044s/rec
195
- ep1 step 1920/3144 loss 0.470 kl 0.000 anchor 0.000 0.044s/rec
196
- ep1 step 1930/3144 loss 0.438 kl 0.000 anchor 0.000 0.044s/rec
197
- ep1 step 1940/3144 loss 0.431 kl 0.000 anchor 0.000 0.044s/rec
198
- ep1 step 1950/3144 loss 0.819 kl 0.000 anchor 0.000 0.044s/rec
199
- ep1 step 1960/3144 loss 0.525 kl 0.000 anchor 0.000 0.044s/rec
200
- ep1 step 1970/3144 loss 0.360 kl 0.000 anchor 0.000 0.044s/rec
201
- ep1 step 1980/3144 loss 0.377 kl 0.000 anchor 0.000 0.044s/rec
202
- ep1 step 1990/3144 loss 0.380 kl 0.000 anchor 0.000 0.044s/rec
203
- ep1 step 2000/3144 loss 0.269 kl 0.000 anchor 0.000 0.044s/rec
204
- ep1 step 2010/3144 loss 0.759 kl 0.000 anchor 0.000 0.043s/rec
205
- ep1 step 2020/3144 loss 0.640 kl 0.000 anchor 0.000 0.043s/rec
206
- ep1 step 2030/3144 loss 0.539 kl 0.000 anchor 0.000 0.043s/rec
207
- ep1 step 2040/3144 loss 0.324 kl 0.000 anchor 0.000 0.043s/rec
208
- ep1 step 2050/3144 loss 0.239 kl 0.000 anchor 0.000 0.043s/rec
209
- ep1 step 2060/3144 loss 0.484 kl 0.000 anchor 0.000 0.043s/rec
210
- ep1 step 2070/3144 loss 0.231 kl 0.000 anchor 0.000 0.043s/rec
211
- ep1 step 2080/3144 loss 0.365 kl 0.000 anchor 0.000 0.043s/rec
212
- ep1 step 2090/3144 loss 0.206 kl 0.000 anchor 0.000 0.043s/rec
213
- ep1 step 2100/3144 loss 0.512 kl 0.000 anchor 0.000 0.043s/rec
214
- ep1 step 2110/3144 loss 0.314 kl 0.000 anchor 0.000 0.043s/rec
215
- ep1 step 2120/3144 loss 0.365 kl 0.000 anchor 0.000 0.043s/rec
216
- ep1 step 2130/3144 loss 0.279 kl 0.000 anchor 0.000 0.043s/rec
217
- ep1 step 2140/3144 loss 0.364 kl 0.000 anchor 0.000 0.043s/rec
218
- ep1 step 2150/3144 loss 0.480 kl 0.000 anchor 0.000 0.043s/rec
219
- ep1 step 2160/3144 loss 0.417 kl 0.000 anchor 0.000 0.043s/rec
220
- ep1 step 2170/3144 loss 0.198 kl 0.000 anchor 0.000 0.043s/rec
221
- ep1 step 2180/3144 loss 0.258 kl 0.000 anchor 0.000 0.043s/rec
222
- ep1 step 2190/3144 loss 0.406 kl 0.000 anchor 0.000 0.043s/rec
223
- ep1 step 2200/3144 loss 0.400 kl 0.000 anchor 0.000 0.043s/rec
224
- ep1 step 2210/3144 loss 0.270 kl 0.000 anchor 0.000 0.042s/rec
225
- ep1 step 2220/3144 loss 0.315 kl 0.000 anchor 0.000 0.042s/rec
226
- ep1 step 2230/3144 loss 0.312 kl 0.000 anchor 0.000 0.042s/rec
227
- ep1 step 2240/3144 loss 0.564 kl 0.000 anchor 0.000 0.042s/rec
228
- ep1 step 2250/3144 loss 0.321 kl 0.000 anchor 0.000 0.042s/rec
229
- ep1 step 2260/3144 loss 0.364 kl 0.000 anchor 0.000 0.042s/rec
230
- ep1 step 2270/3144 loss 0.433 kl 0.000 anchor 0.000 0.042s/rec
231
- ep1 step 2280/3144 loss 0.484 kl 0.000 anchor 0.000 0.042s/rec
232
- ep1 step 2290/3144 loss 0.236 kl 0.000 anchor 0.000 0.042s/rec
233
- ep1 step 2300/3144 loss 0.289 kl 0.000 anchor 0.000 0.042s/rec
234
- ep1 step 2310/3144 loss 0.336 kl 0.000 anchor 0.000 0.042s/rec
235
- ep1 step 2320/3144 loss 0.286 kl 0.000 anchor 0.000 0.042s/rec
236
- ep1 step 2330/3144 loss 0.384 kl 0.000 anchor 0.000 0.042s/rec
237
- ep1 step 2340/3144 loss 0.308 kl 0.000 anchor 0.000 0.042s/rec
238
- ep1 step 2350/3144 loss 0.423 kl 0.000 anchor 0.000 0.042s/rec
239
- ep1 step 2360/3144 loss 0.442 kl 0.000 anchor 0.000 0.042s/rec
240
- ep1 step 2370/3144 loss 0.455 kl 0.000 anchor 0.000 0.042s/rec
241
- ep1 step 2380/3144 loss 0.432 kl 0.000 anchor 0.000 0.042s/rec
242
- ep1 step 2390/3144 loss 0.226 kl 0.000 anchor 0.000 0.042s/rec
243
- ep1 step 2400/3144 loss 0.236 kl 0.000 anchor 0.000 0.042s/rec
244
- ep1 step 2410/3144 loss 0.330 kl 0.000 anchor 0.000 0.042s/rec
245
- ep1 step 2420/3144 loss 0.444 kl 0.000 anchor 0.000 0.042s/rec
246
- ep1 step 2430/3144 loss 0.421 kl 0.000 anchor 0.000 0.041s/rec
247
- ep1 step 2440/3144 loss 0.304 kl 0.000 anchor 0.000 0.041s/rec
248
- ep1 step 2450/3144 loss 0.342 kl 0.000 anchor 0.000 0.041s/rec
249
- ep1 step 2460/3144 loss 0.307 kl 0.000 anchor 0.000 0.041s/rec
250
- ep1 step 2470/3144 loss 0.424 kl 0.000 anchor 0.000 0.041s/rec
251
- ep1 step 2480/3144 loss 0.297 kl 0.000 anchor 0.000 0.041s/rec
252
- ep1 step 2490/3144 loss 0.490 kl 0.000 anchor 0.000 0.041s/rec
253
- ep1 step 2500/3144 loss 0.440 kl 0.000 anchor 0.000 0.041s/rec
254
- ep1 step 2510/3144 loss 0.437 kl 0.000 anchor 0.000 0.041s/rec
255
- ep1 step 2520/3144 loss 0.300 kl 0.000 anchor 0.000 0.041s/rec
256
- ep1 step 2530/3144 loss 0.344 kl 0.000 anchor 0.000 0.041s/rec
257
- ep1 step 2540/3144 loss 0.255 kl 0.000 anchor 0.000 0.041s/rec
258
- ep1 step 2550/3144 loss 0.391 kl 0.000 anchor 0.000 0.041s/rec
259
- ep1 step 2560/3144 loss 0.351 kl 0.000 anchor 0.000 0.041s/rec
260
- ep1 step 2570/3144 loss 0.435 kl 0.000 anchor 0.000 0.041s/rec
261
- ep1 step 2580/3144 loss 0.225 kl 0.000 anchor 0.000 0.041s/rec
262
- ep1 step 2590/3144 loss 0.497 kl 0.000 anchor 0.000 0.041s/rec
263
- ep1 step 2600/3144 loss 0.273 kl 0.000 anchor 0.000 0.041s/rec
264
- ep1 step 2610/3144 loss 0.688 kl 0.000 anchor 0.000 0.041s/rec
265
- ep1 step 2620/3144 loss 0.203 kl 0.000 anchor 0.000 0.041s/rec
266
- ep1 step 2630/3144 loss 0.493 kl 0.000 anchor 0.000 0.041s/rec
267
- ep1 step 2640/3144 loss 0.472 kl 0.000 anchor 0.000 0.041s/rec
268
- ep1 step 2650/3144 loss 0.477 kl 0.000 anchor 0.000 0.041s/rec
269
- ep1 step 2660/3144 loss 0.281 kl 0.000 anchor 0.000 0.041s/rec
270
- ep1 step 2670/3144 loss 0.409 kl 0.000 anchor 0.000 0.041s/rec
271
- ep1 step 2680/3144 loss 0.498 kl 0.000 anchor 0.000 0.041s/rec
272
- ep1 step 2690/3144 loss 0.341 kl 0.000 anchor 0.000 0.040s/rec
273
- ep1 step 2700/3144 loss 0.399 kl 0.000 anchor 0.000 0.040s/rec
274
- ep1 step 2710/3144 loss 0.279 kl 0.000 anchor 0.000 0.040s/rec
275
- ep1 step 2720/3144 loss 0.324 kl 0.000 anchor 0.000 0.040s/rec
276
- ep1 step 2730/3144 loss 0.220 kl 0.000 anchor 0.000 0.040s/rec
277
- ep1 step 2740/3144 loss 0.427 kl 0.000 anchor 0.000 0.040s/rec
278
- ep1 step 2750/3144 loss 0.173 kl 0.000 anchor 0.000 0.040s/rec
279
- ep1 step 2760/3144 loss 0.498 kl 0.000 anchor 0.000 0.040s/rec
280
- ep1 step 2770/3144 loss 0.262 kl 0.000 anchor 0.000 0.040s/rec
281
- ep1 step 2780/3144 loss 0.220 kl 0.000 anchor 0.000 0.040s/rec
282
- ep1 step 2790/3144 loss 0.169 kl 0.000 anchor 0.000 0.040s/rec
283
- ep1 step 2800/3144 loss 0.300 kl 0.000 anchor 0.000 0.040s/rec
284
- ep1 step 2810/3144 loss 0.259 kl 0.000 anchor 0.000 0.040s/rec
285
- ep1 step 2820/3144 loss 0.334 kl 0.000 anchor 0.000 0.040s/rec
286
- ep1 step 2830/3144 loss 0.379 kl 0.000 anchor 0.000 0.040s/rec
287
- ep1 step 2840/3144 loss 0.328 kl 0.000 anchor 0.000 0.040s/rec
288
- ep1 step 2850/3144 loss 0.320 kl 0.000 anchor 0.000 0.040s/rec
289
- ep1 step 2860/3144 loss 0.357 kl 0.000 anchor 0.000 0.040s/rec
290
- ep1 step 2870/3144 loss 0.481 kl 0.000 anchor 0.000 0.040s/rec
291
- ep1 step 2880/3144 loss 0.496 kl 0.000 anchor 0.000 0.040s/rec
292
- ep1 step 2890/3144 loss 0.596 kl 0.000 anchor 0.000 0.040s/rec
293
- ep1 step 2900/3144 loss 0.521 kl 0.000 anchor 0.000 0.040s/rec
294
- ep1 step 2910/3144 loss 0.150 kl 0.000 anchor 0.000 0.040s/rec
295
- ep1 step 2920/3144 loss 0.139 kl 0.000 anchor 0.000 0.040s/rec
296
- ep1 step 2930/3144 loss 0.261 kl 0.000 anchor 0.000 0.040s/rec
297
- ep1 step 2940/3144 loss 0.564 kl 0.000 anchor 0.000 0.040s/rec
298
- ep1 step 2950/3144 loss 0.383 kl 0.000 anchor 0.000 0.040s/rec
299
- ep1 step 2960/3144 loss 0.354 kl 0.000 anchor 0.000 0.040s/rec
300
- ep1 step 2970/3144 loss 0.373 kl 0.000 anchor 0.000 0.040s/rec
301
- ep1 step 2980/3144 loss 0.119 kl 0.000 anchor 0.000 0.040s/rec
302
- ep1 step 2990/3144 loss 0.469 kl 0.000 anchor 0.000 0.040s/rec
303
- ep1 step 3000/3144 loss 0.416 kl 0.000 anchor 0.000 0.040s/rec
304
- ep1 step 3010/3144 loss 0.341 kl 0.000 anchor 0.000 0.040s/rec
305
- ep1 step 3020/3144 loss 0.405 kl 0.000 anchor 0.000 0.040s/rec
306
- ep1 step 3030/3144 loss 0.549 kl 0.000 anchor 0.000 0.040s/rec
307
- ep1 step 3040/3144 loss 0.288 kl 0.000 anchor 0.000 0.040s/rec
308
- ep1 step 3050/3144 loss 0.255 kl 0.000 anchor 0.000 0.039s/rec
309
- ep1 step 3060/3144 loss 0.327 kl 0.000 anchor 0.000 0.039s/rec
310
- ep1 step 3070/3144 loss 0.322 kl 0.000 anchor 0.000 0.039s/rec
311
- ep1 step 3080/3144 loss 0.485 kl 0.000 anchor 0.000 0.040s/rec
312
- ep1 step 3090/3144 loss 0.222 kl 0.000 anchor 0.000 0.039s/rec
313
- ep1 step 3100/3144 loss 0.311 kl 0.000 anchor 0.000 0.039s/rec
314
- ep1 step 3110/3144 loss 0.452 kl 0.000 anchor 0.000 0.039s/rec
315
- ep1 step 3120/3144 loss 0.485 kl 0.000 anchor 0.000 0.039s/rec
316
- ep1 step 3130/3144 loss 0.277 kl 0.000 anchor 0.000 0.039s/rec
317
- ep1 step 3140/3144 loss 0.230 kl 0.000 anchor 0.000 0.039s/rec
318
- Warning: You are sending unauthenticated requests to the HF Hub. Please set a HF_TOKEN to enable higher rate limits and faster downloads.
319
- saved /runs/q35-08b/02-trial-2/checkpoint
 
1
+ Warning: You are sending unauthenticated requests to the HF Hub. Please set a HF_TOKEN to enable higher rate limits and faster downloads.
2
+ delta: warm start from /__modal/volumes/vo-kEMu8BkBAIrorAQI6V8f2D/hub/models--jaredpalmer--kev-0.8b/snapshots/c917edefdfd72b3e9ba71455584700acc70595f6: 372 adapter tensors and the pointer head loaded
3
  device=cuda trainable params=11.3M
4
+ replay: 2000 of 12576 suite training records mixed with 1425 from evals/night2/dates_unknowable.jsonl
5
+ 3425 training requests (holdout=[]), questions by type {'noul': 1268, 'score': 1431, 'choice': 1214}
6
  [transformers] `causal_conv1d_fn` is falling back to its reference PyTorch implementation because `causal_conv1d` is not installed. This is correct but much slower; install `causal_conv1d` for the optimized kernel.
7
+ ep0 step 10/429 loss 0.364 kl 0.000 anchor 0.000 1.620s/rec
8
+ ep0 step 20/429 loss 0.487 kl 0.000 anchor 0.000 0.946s/rec
9
+ ep0 step 30/429 loss 0.310 kl 0.000 anchor 0.000 0.650s/rec
10
+ ep0 step 40/429 loss 0.476 kl 0.000 anchor 0.000 0.506s/rec
11
+ ep0 step 50/429 loss 0.393 kl 0.000 anchor 0.000 0.416s/rec
12
+ ep0 step 60/429 loss 0.418 kl 0.000 anchor 0.000 0.346s/rec
13
+ ep0 step 70/429 loss 0.449 kl 0.000 anchor 0.000 0.301s/rec
14
+ ep0 step 80/429 loss 0.461 kl 0.000 anchor 0.000 0.266s/rec
15
+ ep0 step 90/429 loss 0.257 kl 0.000 anchor 0.000 0.240s/rec
16
+ ep0 step 100/429 loss 0.734 kl 0.000 anchor 0.000 0.222s/rec
17
+ ep0 step 110/429 loss 0.427 kl 0.000 anchor 0.000 0.205s/rec
18
+ ep0 step 120/429 loss 0.269 kl 0.000 anchor 0.000 0.190s/rec
19
+ ep0 step 130/429 loss 0.658 kl 0.000 anchor 0.000 0.177s/rec
20
+ ep0 step 140/429 loss 0.192 kl 0.000 anchor 0.000 0.181s/rec
21
+ ep0 step 150/429 loss 0.390 kl 0.000 anchor 0.000 0.171s/rec
22
+ ep0 step 160/429 loss 0.528 kl 0.000 anchor 0.000 0.161s/rec
23
+ ep0 step 170/429 loss 0.385 kl 0.000 anchor 0.000 0.153s/rec
24
+ ep0 step 180/429 loss 0.368 kl 0.000 anchor 0.000 0.146s/rec
25
+ ep0 step 190/429 loss 0.362 kl 0.000 anchor 0.000 0.140s/rec
26
+ ep0 step 200/429 loss 0.365 kl 0.000 anchor 0.000 0.135s/rec
27
+ ep0 step 210/429 loss 0.267 kl 0.000 anchor 0.000 0.130s/rec
28
+ ep0 step 220/429 loss 0.463 kl 0.000 anchor 0.000 0.125s/rec
29
+ ep0 step 230/429 loss 0.235 kl 0.000 anchor 0.000 0.121s/rec
30
+ ep0 step 240/429 loss 0.228 kl 0.000 anchor 0.000 0.117s/rec
31
+ ep0 step 250/429 loss 0.267 kl 0.000 anchor 0.000 0.113s/rec
32
+ ep0 step 260/429 loss 0.424 kl 0.000 anchor 0.000 0.110s/rec
33
+ ep0 step 270/429 loss 0.231 kl 0.000 anchor 0.000 0.107s/rec
34
+ ep0 step 280/429 loss 0.409 kl 0.000 anchor 0.000 0.104s/rec
35
+ ep0 step 290/429 loss 0.401 kl 0.000 anchor 0.000 0.102s/rec
36
+ ep0 step 300/429 loss 0.402 kl 0.000 anchor 0.000 0.099s/rec
37
+ ep0 step 310/429 loss 0.264 kl 0.000 anchor 0.000 0.097s/rec
38
+ ep0 step 320/429 loss 0.337 kl 0.000 anchor 0.000 0.095s/rec
39
+ ep0 step 330/429 loss 0.552 kl 0.000 anchor 0.000 0.093s/rec
40
+ ep0 step 340/429 loss 0.436 kl 0.000 anchor 0.000 0.091s/rec
41
+ ep0 step 350/429 loss 0.309 kl 0.000 anchor 0.000 0.089s/rec
42
+ ep0 step 360/429 loss 0.239 kl 0.000 anchor 0.000 0.088s/rec
43
+ ep0 step 370/429 loss 0.449 kl 0.000 anchor 0.000 0.087s/rec
44
+ ep0 step 380/429 loss 0.207 kl 0.000 anchor 0.000 0.085s/rec
45
+ ep0 step 390/429 loss 0.430 kl 0.000 anchor 0.000 0.084s/rec
46
+ ep0 step 400/429 loss 0.216 kl 0.000 anchor 0.000 0.082s/rec
47
+ ep0 step 410/429 loss 0.219 kl 0.000 anchor 0.000 0.081s/rec
48
+ ep0 step 420/429 loss 0.284 kl 0.000 anchor 0.000 0.080s/rec
49
+ saved /runs/night2-08b-du2/00-trial-0/checkpoint
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
training_config.json CHANGED
@@ -2,8 +2,8 @@
2
  "args": {
3
  "base": "Qwen/Qwen3.5-0.8B-Base",
4
  "n_per_source": 1000,
5
- "epochs": 2,
6
- "lr": 0.0001,
7
  "head_lr": 0.0,
8
  "weight_decay": 0.01,
9
  "lora": 16,
@@ -17,6 +17,7 @@
17
  "device": "cuda",
18
  "batch": 8,
19
  "dtype": "bf16",
 
20
  "checkpointing": 0,
21
  "option_isolation": 0,
22
  "special_embeddings": 0,
@@ -32,11 +33,20 @@
32
  "anchor": "",
33
  "anchor_w": 0.0,
34
  "anchor_sources": "",
35
- "out": "/runs/q35-08b/02-trial-2/checkpoint",
36
- "seed": 2
 
 
 
37
  },
38
  "suite_sha256": "a8f50e481b7d90b97da049e0ff6a01cee2f1ed204aed61a8265af0edbb5514d2",
39
  "base_revision": "dc7cdfe2ee4154fa7e30f5b51ca41bfa40174e68",
 
 
 
 
 
 
40
  "ordinal_objective": "ranked_probability_score",
41
  "holdout": []
42
  }
 
2
  "args": {
3
  "base": "Qwen/Qwen3.5-0.8B-Base",
4
  "n_per_source": 1000,
5
+ "epochs": 1,
6
+ "lr": 4e-05,
7
  "head_lr": 0.0,
8
  "weight_decay": 0.01,
9
  "lora": 16,
 
17
  "device": "cuda",
18
  "batch": 8,
19
  "dtype": "bf16",
20
+ "weights_dtype": "fp32",
21
  "checkpointing": 0,
22
  "option_isolation": 0,
23
  "special_embeddings": 0,
 
33
  "anchor": "",
34
  "anchor_w": 0.0,
35
  "anchor_sources": "",
36
+ "out": "/runs/night2-08b-du2/00-trial-0/checkpoint",
37
+ "data": "evals/night2/dates_unknowable.jsonl",
38
+ "replay": 2000,
39
+ "init_from": "jaredpalmer/kev-0.8b",
40
+ "seed": 1
41
  },
42
  "suite_sha256": "a8f50e481b7d90b97da049e0ff6a01cee2f1ed204aed61a8265af0edbb5514d2",
43
  "base_revision": "dc7cdfe2ee4154fa7e30f5b51ca41bfa40174e68",
44
+ "init_source": {
45
+ "init_from": "jaredpalmer/kev-0.8b",
46
+ "resolved": "/__modal/volumes/vo-kEMu8BkBAIrorAQI6V8f2D/hub/models--jaredpalmer--kev-0.8b/snapshots/c917edefdfd72b3e9ba71455584700acc70595f6",
47
+ "adapter_sha256": "d9fa619fd3b0490122454c386bfa1b53c23850189d1b5c4346962162e2e39a64",
48
+ "head_sha256": "8610dac1c30a64bbcea7715f20a258f4d084d7b862c10f96a1625c732049a32a"
49
+ },
50
  "ordinal_objective": "ranked_probability_score",
51
  "holdout": []
52
  }
training_metrics.json CHANGED
@@ -1,14 +1,14 @@
1
  {
2
- "wall_seconds": 1195.3522906303406,
3
- "records_seen": 30370,
4
- "requested_records": 25152,
5
  "truncated_records": 0,
6
  "rejected_records": 0,
7
- "optimizer_steps": 3144,
8
- "forward_tokens": 5832366,
9
- "peak_device_bytes": 56761799680,
10
  "device": "cuda",
11
  "dtype": "bf16",
12
  "batch": 8,
13
- "peak_rss_bytes": 10316271616
14
  }
 
1
  {
2
+ "wall_seconds": 313.94338822364807,
3
+ "records_seen": 3937,
4
+ "requested_records": 3425,
5
  "truncated_records": 0,
6
  "rejected_records": 0,
7
+ "optimizer_steps": 429,
8
+ "forward_tokens": 640184,
9
+ "peak_device_bytes": 43896273920,
10
  "device": "cuda",
11
  "dtype": "bf16",
12
  "batch": 8,
13
+ "peak_rss_bytes": 10316816384
14
  }