dougalldeepmind/2026-07-31-qwen36-difficult-advice-tulu-lora-40-60-assistant-loss-only
Image-Text-to-Text • Updated • 16
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
n_supervised: int64
n_tokens: int64
spans: list<item: list<item: int64>>
child 0, item: list<item: int64>
child 0, item: int64
row: 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
n_supervised: int64
n_tokens: int64
spans: list<item: list<item: int64>>
child 0, item: list<item: int64>
child 0, item: int64
row: 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 | difficult_advice | 990 | 831 | [
[
700,
4927
]
] |
1 | tulu3 | 374 | 77 | [
[
1376,
1831
]
] |
2 | difficult_advice | 1,088 | 959 | [
[
647,
5377
]
] |
3 | tulu3 | 1,107 | 1,016 | [
[
237,
2736
]
] |
4 | tulu3 | 245 | 185 | [
[
281,
1219
]
] |
5 | difficult_advice | 1,106 | 944 | [
[
762,
5480
]
] |
6 | tulu3 | 79 | 54 | [
[
140,
435
]
] |
7 | tulu3 | 243 | 107 | [
[
721,
1219
]
] |
8 | tulu3 | 58 | 37 | [
[
94,
247
]
] |
9 | difficult_advice | 977 | 838 | [
[
620,
4934
]
] |
10 | difficult_advice | 953 | 801 | [
[
779,
4731
]
] |
11 | tulu3 | 283 | 238 | [
[
80,
541
],
[
665,
1167
]
] |
12 | difficult_advice | 1,033 | 874 | [
[
826,
5487
]
] |
13 | tulu3 | 187 | 130 | [
[
204,
624
]
] |
14 | difficult_advice | 928 | 761 | [
[
695,
4441
]
] |
15 | tulu3 | 1,685 | 1,633 | [
[
89,
3254
],
[
3391,
6334
],
[
6441,
9622
]
] |
16 | tulu3 | 314 | 283 | [
[
105,
780
]
] |
17 | tulu3 | 145 | 111 | [
[
94,
425
],
[
499,
592
]
] |
18 | tulu3 | 367 | 333 | [
[
133,
1705
]
] |
19 | difficult_advice | 1,167 | 1,048 | [
[
613,
5872
]
] |
20 | difficult_advice | 966 | 834 | [
[
686,
5006
]
] |
21 | difficult_advice | 1,104 | 940 | [
[
732,
5596
]
] |
22 | tulu3 | 1,180 | 904 | [
[
1166,
3697
]
] |
23 | tulu3 | 1,802 | 1,424 | [
[
1369,
4194
]
] |
24 | tulu3 | 675 | 609 | [
[
188,
1861
]
] |
25 | difficult_advice | 930 | 797 | [
[
673,
4882
]
] |
26 | tulu3 | 1,247 | 988 | [
[
632,
4466
]
] |
27 | tulu3 | 873 | 844 | [
[
106,
3598
]
] |
28 | tulu3 | 213 | 198 | [
[
90,
1030
]
] |
29 | tulu3 | 374 | 287 | [
[
426,
1705
]
] |
30 | tulu3 | 1,111 | 874 | [
[
1137,
3461
]
] |
31 | difficult_advice | 1,133 | 962 | [
[
802,
5536
]
] |
32 | tulu3 | 1,755 | 1,400 | [
[
1454,
5331
]
] |
33 | difficult_advice | 1,125 | 947 | [
[
771,
5498
]
] |
34 | tulu3 | 291 | 233 | [
[
292,
1579
]
] |
35 | tulu3 | 23 | 6 | [
[
78,
115
]
] |
36 | difficult_advice | 1,071 | 906 | [
[
848,
5820
]
] |
37 | difficult_advice | 1,004 | 876 | [
[
625,
5185
]
] |
38 | tulu3 | 1,145 | 949 | [
[
925,
3217
]
] |
39 | tulu3 | 727 | 585 | [
[
715,
2834
]
] |
40 | difficult_advice | 1,094 | 916 | [
[
832,
5752
]
] |
41 | difficult_advice | 1,020 | 888 | [
[
638,
5322
]
] |
42 | difficult_advice | 992 | 826 | [
[
835,
5334
]
] |
43 | tulu3 | 849 | 441 | [
[
1298,
2684
]
] |
44 | tulu3 | 286 | 231 | [
[
127,
315
],
[
429,
1089
]
] |
45 | difficult_advice | 1,162 | 982 | [
[
934,
6012
]
] |
46 | tulu3 | 1,001 | 680 | [
[
1298,
3284
]
] |
47 | difficult_advice | 1,097 | 973 | [
[
601,
5520
]
] |
48 | difficult_advice | 1,257 | 1,118 | [
[
780,
6704
]
] |
49 | tulu3 | 390 | 282 | [
[
95,
625
],
[
790,
855
]
] |
50 | tulu3 | 259 | 203 | [
[
120,
562
],
[
686,
1015
]
] |
51 | tulu3 | 1,112 | 864 | [
[
943,
3882
]
] |
52 | difficult_advice | 1,219 | 1,057 | [
[
838,
6165
]
] |
53 | difficult_advice | 971 | 840 | [
[
661,
5073
]
] |
54 | tulu3 | 903 | 696 | [
[
585,
3338
]
] |
55 | difficult_advice | 1,084 | 902 | [
[
841,
5439
]
] |
56 | difficult_advice | 890 | 742 | [
[
670,
4507
]
] |
57 | tulu3 | 69 | 52 | [
[
93,
333
]
] |
58 | tulu3 | 985 | 807 | [
[
893,
3363
]
] |
59 | tulu3 | 81 | 53 | [
[
147,
417
]
] |
60 | difficult_advice | 972 | 849 | [
[
555,
4898
]
] |
61 | difficult_advice | 989 | 855 | [
[
632,
5013
]
] |
62 | difficult_advice | 1,154 | 980 | [
[
934,
6210
]
] |
63 | tulu3 | 389 | 270 | [
[
405,
1351
]
] |
64 | tulu3 | 199 | 93 | [
[
567,
1112
]
] |
65 | tulu3 | 1,832 | 1,438 | [
[
1649,
4922
]
] |
66 | tulu3 | 232 | 41 | [
[
1001,
1228
]
] |
67 | difficult_advice | 910 | 796 | [
[
567,
4720
]
] |
68 | tulu3 | 538 | 394 | [
[
606,
2091
]
] |
69 | tulu3 | 382 | 347 | [
[
157,
1587
]
] |
70 | difficult_advice | 1,034 | 870 | [
[
818,
5276
]
] |
71 | difficult_advice | 1,301 | 1,111 | [
[
911,
6634
]
] |
72 | difficult_advice | 960 | 789 | [
[
819,
4893
]
] |
73 | tulu3 | 205 | 190 | [
[
89,
1134
]
] |
74 | tulu3 | 229 | 127 | [
[
540,
1286
]
] |
75 | tulu3 | 257 | 77 | [
[
1002,
1455
]
] |
76 | tulu3 | 250 | 228 | [
[
119,
1242
]
] |
77 | difficult_advice | 1,020 | 898 | [
[
608,
5349
]
] |
78 | tulu3 | 241 | 185 | [
[
309,
1144
]
] |
79 | difficult_advice | 1,144 | 984 | [
[
741,
5858
]
] |
80 | difficult_advice | 1,055 | 928 | [
[
656,
5492
]
] |
81 | difficult_advice | 1,050 | 911 | [
[
756,
5740
]
] |
82 | tulu3 | 113 | 91 | [
[
112,
603
]
] |
83 | tulu3 | 593 | 555 | [
[
120,
1521
]
] |
84 | difficult_advice | 917 | 762 | [
[
780,
4801
]
] |
85 | tulu3 | 214 | 121 | [
[
496,
1106
]
] |
86 | tulu3 | 476 | 392 | [
[
250,
1258
]
] |
87 | tulu3 | 77 | 18 | [
[
239,
324
]
] |
88 | difficult_advice | 1,157 | 1,036 | [
[
636,
5896
]
] |
89 | tulu3 | 321 | 252 | [
[
305,
1293
]
] |
90 | tulu3 | 804 | 711 | [
[
183,
1764
]
] |
91 | tulu3 | 1,429 | 440 | [
[
4391,
6711
]
] |
92 | tulu3 | 962 | 116 | [
[
4195,
4804
]
] |
93 | difficult_advice | 959 | 825 | [
[
700,
5129
]
] |
94 | tulu3 | 265 | 194 | [
[
151,
673
]
] |
95 | tulu3 | 1,574 | 1,227 | [
[
1680,
5492
]
] |
96 | tulu3 | 527 | 476 | [
[
152,
1310
]
] |
97 | tulu3 | 816 | 793 | [
[
111,
3109
]
] |
98 | tulu3 | 194 | 161 | [
[
135,
832
]
] |
99 | tulu3 | 156 | 105 | [
[
247,
614
]
] |
End of preview.
Qwen3.6-27B SFT mixture — 40-60_assistant_loss_only
40% difficult-advice / 60% TULU3 replay, by token. Built for training with loss on assistant tokens only.
mixture.jsonl is byte-identical (md5 88f39a3d01e59ba9d592b26c1705c57f, 1,982 rows) to the mixture used by the
full-token arm
…-tulu-lora-40-60, so the loss mask is the only
difference between the two runs.
| Source | Rows | Tokens | Share | Supervised |
|---|---|---|---|---|
| difficult-advice | 580 | 597,013 | 40.0% | 85.48% |
| TULU3 replay | 1,402 | 896,346 | 60.0% | 77.98% |
| Total | 1,982 | 1,493,359 | 80.98% |
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, plus token counts |
stats.json |
the table above, machine-readable |
Reasoning traces
| Data | Renders as |
|---|---|
| difficult-advice (580/580 rows) | <think>real reasoning</think> |
| TULU3 replay (0/1,402 rows) | no <think> block at all |
Zero rows carry an empty <think></think> (asserted at build time). That distinction is the
point: an empty think block is Qwen3.6's explicit do-not-deliberate marker and trains a model to
stop reasoning, whereas absent markup says nothing either way. The builder gets there by
appending a throwaway user turn — which demotes the assistant turn from final to historical, so the
template takes its no-think branch — then cutting that turn back off.
Re-rendering these conversations from messages will not reproduce the training data.
What "assistant_loss_only" means
Every token outside an assistant turn is -100 and contributes no loss. A supervised span starts
immediately after the <|im_start|>assistant\n header — which the model is given at inference and
never has to produce — and ends after the closing <|im_end|>, which it must produce in order to
stop.
<|im_start|> MASKED
assistant MASKED
\n MASKED
<think> LOSS <- supervision starts at the first generated token
assistant_spans.jsonl records those spans as character offsets into text, so the mask is
reproducible without our code.
TRL's assistant_only_loss flag cannot do this on Qwen3.6. It requires {% generation %}
markers the template lacks, and it re-renders from messages, discarding the think-block
convention above. The spans are derived from the rendered text via the fast tokenizer's offset
mapping instead.
Provenance
src/experiments/build_mixture.py with configs/mixture_qwen36_40_60.yaml,
seed 0, from allenai/tulu-3-sft-mixture
and matboz/difficult-advice-qwen3. Replay conversations over 2,048 Qwen3.6 tokens were dropped
rather than truncated.
- Downloads last month
- 154
Models trained or fine-tuned on dougalldeepmind/2026-07-31-qwen36-sft-mixture-40-60-assistant-loss-only