LASR-Callum/2026-07-31-qwen36-difficult-advice-tulu-lora-80-20-empty-think-tags
Image-Text-to-Text • Updated • 19
Datasets:
Dataset Preview
The full dataset viewer is not available (click to read why). Only showing a preview of the rows.
The dataset generation failed
Error code: DatasetGenerationError
Exception: CastError
Message: Couldn't cast
text: string
source: string
spans: list<item: list<item: int64>>
child 0, item: list<item: int64>
child 0, item: int64
row: int64
n_tokens: int64
n_supervised: int64
to
{'row': Value('int64'), 'source': Value('string'), 'n_tokens': Value('int64'), 'n_supervised': Value('int64'), 'spans': List(List(Value('int64')))}
because column names don't match
Traceback: Traceback (most recent call last):
File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1827, in _prepare_split_single
for key, table in generator:
^^^^^^^^^
File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 613, in wrapped
for item in generator(*args, **kwargs):
~~~~~~~~~^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 343, in _generate_tables
self._cast_table(pa_table, json_field_paths=json_field_paths),
~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 132, in _cast_table
pa_table = table_cast(pa_table, self.info.features.arrow_schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2378, in table_cast
return cast_table_to_schema(table, schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2306, in cast_table_to_schema
raise CastError(
...<3 lines>...
)
datasets.table.CastError: Couldn't cast
text: string
source: string
spans: list<item: list<item: int64>>
child 0, item: list<item: int64>
child 0, item: int64
row: int64
n_tokens: int64
n_supervised: int64
to
{'row': Value('int64'), 'source': Value('string'), 'n_tokens': Value('int64'), 'n_supervised': Value('int64'), 'spans': List(List(Value('int64')))}
because column names don't match
The above exception was the direct cause of the following exception:
Traceback (most recent call last):
File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 1369, in compute_config_parquet_and_info_response
parquet_operations, partial, estimated_dataset_info = stream_convert_to_parquet(
~~~~~~~~~~~~~~~~~~~~~~~~~^
builder, max_dataset_size_bytes=max_dataset_size_bytes
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
)
^
File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 948, in stream_convert_to_parquet
builder._prepare_split(split_generator=splits_generators[split], file_format="parquet")
~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1694, in _prepare_split
for job_id, done, content in self._prepare_split_single(
~~~~~~~~~~~~~~~~~~~~~~~~~~^
gen_kwargs=gen_kwargs, job_id=job_id, **_prepare_split_args
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
):
^
File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1880, in _prepare_split_single
raise DatasetGenerationError("An error occurred while generating the dataset") from e
datasets.exceptions.DatasetGenerationError: An error occurred while generating the datasetNeed help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
row int64 | source string | n_tokens int64 | n_supervised int64 | spans list |
|---|---|---|---|---|
0 | tulu3 | 391 | 359 | [
[
172,
1856
]
] |
1 | difficult_advice | 910 | 796 | [
[
567,
4720
]
] |
2 | tulu3 | 261 | 77 | [
[
1021,
1474
]
] |
3 | tulu3 | 449 | 304 | [
[
563,
1741
]
] |
4 | tulu3 | 497 | 262 | [
[
1074,
2069
]
] |
5 | tulu3 | 84 | 52 | [
[
155,
324
]
] |
6 | tulu3 | 1,858 | 1,605 | [
[
1020,
4686
]
] |
7 | difficult_advice | 1,083 | 951 | [
[
666,
5528
]
] |
8 | difficult_advice | 883 | 755 | [
[
656,
4830
]
] |
9 | difficult_advice | 1,020 | 898 | [
[
608,
5349
]
] |
10 | tulu3 | 1,986 | 1,671 | [
[
1294,
5257
]
] |
11 | tulu3 | 131 | 82 | [
[
186,
469
]
] |
12 | tulu3 | 683 | 606 | [
[
330,
2509
]
] |
13 | tulu3 | 530 | 460 | [
[
169,
2037
],
[
2243,
3012
]
] |
14 | tulu3 | 624 | 605 | [
[
95,
2271
]
] |
15 | difficult_advice | 966 | 809 | [
[
745,
5003
]
] |
16 | tulu3 | 662 | 542 | [
[
483,
2486
]
] |
17 | tulu3 | 821 | 454 | [
[
1090,
2941
]
] |
18 | difficult_advice | 959 | 825 | [
[
700,
5129
]
] |
19 | tulu3 | 23 | 3 | [
[
76,
87
]
] |
20 | difficult_advice | 949 | 822 | [
[
608,
4791
]
] |
21 | tulu3 | 795 | 745 | [
[
210,
3031
]
] |
22 | tulu3 | 596 | 522 | [
[
257,
2023
],
[
2144,
2944
]
] |
23 | difficult_advice | 1,094 | 916 | [
[
832,
5752
]
] |
24 | tulu3 | 210 | 107 | [
[
567,
1127
]
] |
25 | tulu3 | 756 | 568 | [
[
845,
2372
]
] |
26 | tulu3 | 1,367 | 695 | [
[
860,
1796
]
] |
27 | difficult_advice | 1,013 | 856 | [
[
718,
5252
]
] |
28 | tulu3 | 240 | 218 | [
[
124,
995
]
] |
29 | tulu3 | 301 | 222 | [
[
197,
643
],
[
736,
759
]
] |
30 | tulu3 | 1,202 | 803 | [
[
1717,
4357
]
] |
31 | tulu3 | 546 | 495 | [
[
199,
1875
]
] |
32 | tulu3 | 558 | 492 | [
[
379,
2505
]
] |
33 | tulu3 | 417 | 379 | [
[
200,
1902
]
] |
34 | tulu3 | 1,774 | 1,477 | [
[
1119,
4273
]
] |
35 | tulu3 | 1,281 | 961 | [
[
1500,
3902
]
] |
36 | tulu3 | 337 | 289 | [
[
255,
1238
]
] |
37 | tulu3 | 859 | 649 | [
[
798,
3367
]
] |
38 | tulu3 | 146 | 113 | [
[
175,
742
]
] |
39 | tulu3 | 1,308 | 1,059 | [
[
1086,
3795
]
] |
40 | tulu3 | 798 | 715 | [
[
153,
1655
],
[
1781,
2528
],
[
2651,
3750
]
] |
41 | tulu3 | 231 | 69 | [
[
866,
1232
]
] |
42 | tulu3 | 473 | 82 | [
[
2041,
2457
]
] |
43 | tulu3 | 739 | 636 | [
[
404,
2786
]
] |
44 | tulu3 | 90 | 62 | [
[
144,
444
]
] |
45 | tulu3 | 236 | 132 | [
[
590,
1353
]
] |
46 | tulu3 | 1,861 | 1,542 | [
[
1191,
4414
]
] |
47 | tulu3 | 205 | 67 | [
[
792,
1162
]
] |
48 | tulu3 | 297 | 246 | [
[
263,
1480
]
] |
49 | tulu3 | 383 | 104 | [
[
1422,
1980
]
] |
50 | difficult_advice | 980 | 840 | [
[
657,
5091
]
] |
51 | tulu3 | 115 | 87 | [
[
126,
523
]
] |
52 | tulu3 | 339 | 286 | [
[
104,
177
],
[
349,
1679
]
] |
53 | tulu3 | 1,638 | 1,511 | [
[
159,
1205
],
[
1277,
2215
],
[
2363,
3405
]
] |
54 | tulu3 | 79 | 53 | [
[
115,
310
]
] |
55 | tulu3 | 301 | 273 | [
[
123,
1395
]
] |
56 | tulu3 | 1,197 | 1,076 | [
[
569,
3353
]
] |
57 | tulu3 | 338 | 242 | [
[
367,
1317
]
] |
58 | tulu3 | 248 | 45 | [
[
827,
1017
]
] |
59 | tulu3 | 994 | 761 | [
[
969,
2720
]
] |
60 | difficult_advice | 1,088 | 959 | [
[
647,
5377
]
] |
61 | tulu3 | 322 | 209 | [
[
81,
412
],
[
800,
1473
]
] |
62 | difficult_advice | 982 | 845 | [
[
655,
5073
]
] |
63 | tulu3 | 795 | 530 | [
[
1055,
2985
]
] |
64 | tulu3 | 537 | 214 | [
[
1664,
2740
]
] |
65 | difficult_advice | 1,038 | 878 | [
[
808,
5450
]
] |
66 | tulu3 | 160 | 105 | [
[
266,
633
]
] |
67 | tulu3 | 231 | 98 | [
[
711,
1305
]
] |
68 | tulu3 | 320 | 282 | [
[
171,
1410
]
] |
69 | tulu3 | 484 | 330 | [
[
831,
2480
]
] |
70 | tulu3 | 1,141 | 962 | [
[
844,
2795
]
] |
71 | difficult_advice | 890 | 742 | [
[
670,
4507
]
] |
72 | difficult_advice | 1,144 | 984 | [
[
741,
5858
]
] |
73 | difficult_advice | 1,084 | 948 | [
[
716,
5757
]
] |
74 | tulu3 | 198 | 89 | [
[
568,
1052
]
] |
75 | tulu3 | 868 | 789 | [
[
369,
4230
]
] |
76 | tulu3 | 415 | 271 | [
[
581,
1875
]
] |
77 | tulu3 | 797 | 463 | [
[
1441,
3092
]
] |
78 | tulu3 | 34 | 12 | [
[
99,
156
]
] |
79 | tulu3 | 1,142 | 931 | [
[
920,
4023
]
] |
80 | tulu3 | 631 | 438 | [
[
742,
2294
]
] |
81 | difficult_advice | 1,033 | 874 | [
[
826,
5487
]
] |
82 | difficult_advice | 930 | 797 | [
[
673,
4882
]
] |
83 | tulu3 | 1,306 | 1,047 | [
[
992,
3476
]
] |
84 | tulu3 | 871 | 815 | [
[
243,
3689
]
] |
85 | tulu3 | 1,230 | 861 | [
[
1872,
4124
]
] |
86 | tulu3 | 394 | 317 | [
[
356,
1585
]
] |
87 | tulu3 | 502 | 296 | [
[
855,
2148
]
] |
88 | difficult_advice | 1,065 | 934 | [
[
613,
5338
]
] |
89 | tulu3 | 661 | 559 | [
[
323,
2572
]
] |
90 | tulu3 | 407 | 327 | [
[
392,
2032
]
] |
91 | tulu3 | 435 | 326 | [
[
489,
2206
]
] |
92 | tulu3 | 298 | 144 | [
[
546,
1144
]
] |
93 | tulu3 | 1,078 | 899 | [
[
456,
2591
]
] |
94 | tulu3 | 1,107 | 640 | [
[
1718,
4055
]
] |
95 | tulu3 | 143 | 121 | [
[
119,
812
]
] |
96 | difficult_advice | 1,054 | 908 | [
[
692,
5270
]
] |
97 | tulu3 | 694 | 622 | [
[
301,
2344
]
] |
98 | tulu3 | 679 | 609 | [
[
207,
1880
]
] |
99 | tulu3 | 498 | 279 | [
[
877,
2064
]
] |
End of preview.
Qwen3.6-27B SFT mixture — 80_20_empty_think_tags
The 20% difficult-advice / 80% TULU3 mixture, with Qwen3.6's empty think marker added to the
replay rows and excluded from the loss. Built for the adapter
qwen3.6-27b-difficult-advice-tulu-lora-80_20_empty_think_tags.
Derived from the 20/80 mixture (md5 7d7da21c632ed31f541f063f507a522f) used by
…-tulu-lora-20-80
and …-20-80-assistant_loss_only. Same 2,169 rows, same 291/1,878
split, same seed. This file: md5 d3d8efa8f483c68eb28ece42839e48dc.
The marker
Every TULU3 replay row carries <think>\n\n</think>\n\n on its final assistant turn --
Qwen3.6's explicit non-thinking marker, placed exactly where apply_chat_template puts it (the
template emits it only on the final turn, never on historical ones; the insertion is asserted to
reproduce the template byte-for-byte before any data is touched).
Those marker tokens are masked out of the loss. The model is conditioned on the marker --
which is how Qwen3.6 injects it as a prefill in non-thinking mode -- but is never trained to
emit it, since learning to emit an empty think block is the documented reasoning-collapse
pattern. Difficult-advice rows are untouched and their real <think> traces stay fully supervised.
<|im_start|> MASKED
assistant MASKED
<think> MASKED <- marker: context, not a target
</think> MASKED
Pre LOSS <- supervision starts at the answer
| Rows | Tokens | Marker | Supervised | |
|---|---|---|---|---|
| difficult-advice | 291 | 299,455 | 0 | 85.45% |
| TULU3 replay | 1,878 | 1,202,056 | 1,878 | 77.51% |
| Total | 2,169 | 1,501,511 | 1,878 | 79.09% |
1,187,560 supervised tokens, versus 1,187,563 in the plain assistant-only 20/80 arm -- the
supervised set is effectively identical, so the marker's presence as context is the only variable.
The 3-token gap is one row (index 1302) that sat at exactly 2,048 tokens and now reaches 2,052,
truncating its trailing <|im_end|>. Left as-is so max_seq_len stays comparable across arms.
Files
| File | What it is |
|---|---|
mixture.jsonl |
the training input: text (pre-rendered) + source |
assistant_spans.jsonl |
per row, the character spans that carried loss (marker already excluded) |
stats.json |
the table above, machine-readable |
Re-rendering from messages will not reproduce these strings.
Provenance
src/experiments/add_empty_think.py applied to output/mixture_qwen36/20260728_152610/mixture.jsonl,
which came from build_mixture.py (configs/mixture_qwen36.yaml, seed 0) over
allenai/tulu-3-sft-mixture
and matboz/difficult-advice-qwen3.
- Downloads last month
- 187
Models trained or fine-tuned on LASR-Callum/2026-07-31-qwen36-sft-mixture-80-20-empty-think-tags