Datasets:

VietNguyenHoang
/

fq-select-math-distillation

Error code:   DatasetGenerationError
Exception:    CastError
Message:      Couldn't cast
benchmarks: struct<math500: struct<acc@4: double, pass@4: double, n_questions: int64, n_generations: int64, mean (... 4197 chars omitted)
  child 0, math500: struct<acc@4: double, pass@4: double, n_questions: int64, n_generations: int64, mean_new_tokens: dou (... 558 chars omitted)
      child 0, acc@4: double
      child 1, pass@4: double
      child 2, n_questions: int64
      child 3, n_generations: int64
      child 4, mean_new_tokens: double
      child 5, stopped_share: double
      child 6, hit_cap_share: double
      child 7, no_boxed_share: double
      child 8, by_subset: struct<Algebra: struct<acc@4: double, pass@4: double>, Counting & Probability: struct<acc@4: double, (... 284 chars omitted)
          child 0, Algebra: struct<acc@4: double, pass@4: double>
              child 0, acc@4: double
              child 1, pass@4: double
          child 1, Counting & Probability: struct<acc@4: double, pass@4: double>
              child 0, acc@4: double
              child 1, pass@4: double
          child 2, Geometry: struct<acc@4: double, pass@4: double>
              child 0, acc@4: double
              child 1, pass@4: double
          child 3, Intermediate Algebra: struct<acc@4: double, pass@4: double>
              child 0, acc@4: double
              child 1, pass@4: double
          child 4, Number Theory: struct<acc@4: double, pass@4: double>
              child 0, acc@4: double
              child 1, pass@4: double
          child 5, Prealgebra: struct<acc@4: d
...
 double
              child 1, pass@4: double
      child 9, seconds: double
      child 10, tokens_per_second: double
      child 11, oom_splits: int64
      child 12, stages: list<item: struct<cap: int64, batch: int64, rows: int64, steps: int64, tokens: int64, seconds: doubl (... 24 chars omitted)
          child 0, item: struct<cap: int64, batch: int64, rows: int64, steps: int64, tokens: int64, seconds: double, ms_per_s (... 12 chars omitted)
              child 0, cap: int64
              child 1, batch: int64
              child 2, rows: int64
              child 3, steps: int64
              child 4, tokens: int64
              child 5, seconds: double
              child 6, ms_per_step: double
      child 13, finished: timestamp[s]
tag: string
seed: int64
student: string
adapter: string
train_commit: string
eval_commit: string
stop_tokens: list<item: string>
  child 0, item: string
settings: struct<n_samples: int64, temperature: double, top_p: double, top_k: int64, max_new_tokens: int64, se (... 108 chars omitted)
  child 0, n_samples: int64
  child 1, temperature: double
  child 2, top_p: double
  child 3, top_k: int64
  child 4, max_new_tokens: int64
  child 5, seed: int64
  child 6, stage_tokens: list<item: int64>
      child 0, item: int64
  child 7, gen_batch: list<item: int64>
      child 0, item: int64
  child 8, chunk: int64
  child 9, bbh_per_task: int64
peak_reserved_gb: double
n_tokens: int64
correct: bool
sample: int64
qid: string
stopped: bool
pred: string
to
{'qid': Value('string'), 'sample': Value('int64'), 'correct': Value('bool'), 'n_tokens': Value('int64'), 'stopped': Value('bool'), 'pred': Value('string')}
because column names don't match
Traceback:    Traceback (most recent call last):
                File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1835, in _prepare_split_single
                  for key, table in generator:
                                    ^^^^^^^^^
                File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 621, in wrapped
                  for item in generator(*args, **kwargs):
                              ~~~~~~~~~^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 361, 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 2381, in table_cast
                  return cast_table_to_schema(table, schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2309, in cast_table_to_schema
                  raise CastError(
                  ...<3 lines>...
                  )
              datasets.table.CastError: Couldn't cast
              benchmarks: struct<math500: struct<acc@4: double, pass@4: double, n_questions: int64, n_generations: int64, mean (... 4197 chars omitted)
                child 0, math500: struct<acc@4: double, pass@4: double, n_questions: int64, n_generations: int64, mean_new_tokens: dou (... 558 chars omitted)
                    child 0, acc@4: double
                    child 1, pass@4: double
                    child 2, n_questions: int64
                    child 3, n_generations: int64
                    child 4, mean_new_tokens: double
                    child 5, stopped_share: double
                    child 6, hit_cap_share: double
                    child 7, no_boxed_share: double
                    child 8, by_subset: struct<Algebra: struct<acc@4: double, pass@4: double>, Counting & Probability: struct<acc@4: double, (... 284 chars omitted)
                        child 0, Algebra: struct<acc@4: double, pass@4: double>
                            child 0, acc@4: double
                            child 1, pass@4: double
                        child 1, Counting & Probability: struct<acc@4: double, pass@4: double>
                            child 0, acc@4: double
                            child 1, pass@4: double
                        child 2, Geometry: struct<acc@4: double, pass@4: double>
                            child 0, acc@4: double
                            child 1, pass@4: double
                        child 3, Intermediate Algebra: struct<acc@4: double, pass@4: double>
                            child 0, acc@4: double
                            child 1, pass@4: double
                        child 4, Number Theory: struct<acc@4: double, pass@4: double>
                            child 0, acc@4: double
                            child 1, pass@4: double
                        child 5, Prealgebra: struct<acc@4: d
              ...
               double
                            child 1, pass@4: double
                    child 9, seconds: double
                    child 10, tokens_per_second: double
                    child 11, oom_splits: int64
                    child 12, stages: list<item: struct<cap: int64, batch: int64, rows: int64, steps: int64, tokens: int64, seconds: doubl (... 24 chars omitted)
                        child 0, item: struct<cap: int64, batch: int64, rows: int64, steps: int64, tokens: int64, seconds: double, ms_per_s (... 12 chars omitted)
                            child 0, cap: int64
                            child 1, batch: int64
                            child 2, rows: int64
                            child 3, steps: int64
                            child 4, tokens: int64
                            child 5, seconds: double
                            child 6, ms_per_step: double
                    child 13, finished: timestamp[s]
              tag: string
              seed: int64
              student: string
              adapter: string
              train_commit: string
              eval_commit: string
              stop_tokens: list<item: string>
                child 0, item: string
              settings: struct<n_samples: int64, temperature: double, top_p: double, top_k: int64, max_new_tokens: int64, se (... 108 chars omitted)
                child 0, n_samples: int64
                child 1, temperature: double
                child 2, top_p: double
                child 3, top_k: int64
                child 4, max_new_tokens: int64
                child 5, seed: int64
                child 6, stage_tokens: list<item: int64>
                    child 0, item: int64
                child 7, gen_batch: list<item: int64>
                    child 0, item: int64
                child 8, chunk: int64
                child 9, bbh_per_task: int64
              peak_reserved_gb: double
              n_tokens: int64
              correct: bool
              sample: int64
              qid: string
              stopped: bool
              pred: string
              to
              {'qid': Value('string'), 'sample': Value('int64'), 'correct': Value('bool'), 'n_tokens': Value('int64'), 'stopped': Value('bool'), 'pred': Value('string')}
              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 1379, 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 956, 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 1702, 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 1887, in _prepare_split_single
                  raise DatasetGenerationError("An error occurred while generating the dataset") from e
              datasets.exceptions.DatasetGenerationError: An error occurred while generating the dataset

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

qid string	sample int64	correct bool	n_tokens int64	stopped bool	pred string
aime24_00000	0	false	2,467	true	51
aime24_00000	1	false	1,047	true	13
aime24_00000	2	false	3,036	true	7
aime24_00000	3	false	3,072	false	null
aime24_00001	0	false	1,109	true	221
aime24_00001	1	false	623	true	11
aime24_00001	2	false	625	true	5
aime24_00001	3	false	1,703	true	115
aime24_00002	0	false	673	true	10
aime24_00002	1	false	553	true	181
aime24_00002	2	false	1,199	true	8
aime24_00002	3	false	1,826	true	2
aime24_00003	0	false	859	true	18
aime24_00003	1	false	581	true	405
aime24_00003	2	false	675	true	1620
aime24_00003	3	false	758	true	1013
aime24_00004	0	false	420	true	1025
aime24_00004	1	false	3,072	false	null
aime24_00004	2	false	491	true	538
aime24_00004	3	false	3,072	false	null
aime24_00005	0	false	607	true	2
aime24_00005	1	false	844	true	8
aime24_00005	2	false	3,072	false	null
aime24_00005	3	false	831	true	6
aime24_00006	0	false	900	true	133
aime24_00006	1	false	558	true	499
aime24_00006	2	false	3,072	false	null
aime24_00006	3	false	3,072	false	null
aime24_00007	0	false	2,640	true	1
aime24_00007	1	false	3,072	false	null
aime24_00007	2	false	1,303	true	12
aime24_00007	3	false	756	true	0
aime24_00008	0	true	739	true	25
aime24_00008	1	true	2,692	true	25
aime24_00008	2	false	650	true	78125
aime24_00008	3	false	3,072	false	null
aime24_00009	0	false	548	true	21
aime24_00009	1	false	567	true	2016
aime24_00009	2	false	463	true	2024
aime24_00009	3	false	755	true	1012024
aime24_00010	0	false	1,405	true	324
aime24_00010	1	false	1,398	true	441
aime24_00010	2	false	1,336	true	162
aime24_00010	3	false	1,081	true	684
aime24_00011	0	false	832	true	0
aime24_00011	1	false	930	true	120
aime24_00011	2	false	551	true	8
aime24_00011	3	false	552	true	4
aime24_00012	0	false	3,072	false	null
aime24_00012	1	false	3,072	false	null
aime24_00012	2	false	859	true	4258
aime24_00012	3	false	3,072	false	null
aime24_00013	0	false	476	true	4994
aime24_00013	1	false	730	true	2016
aime24_00013	2	false	507	true	3521
aime24_00013	3	false	2,674	true	1006
aime24_00014	0	false	817	true	171532800
aime24_00014	1	false	460	true	3003
aime24_00014	2	false	655	true	11440
aime24_00014	3	false	498	true	8008
aime24_00015	0	false	710	true	626
aime24_00015	1	false	3,072	false	null
aime24_00015	2	false	1,025	true	12
aime24_00015	3	false	3,072	false	null
aime24_00016	0	false	3,072	false	null
aime24_00016	1	false	555	true	533
aime24_00016	2	false	776	true	10
aime24_00016	3	false	1,406	true	533
aime24_00017	0	false	287	true	36
aime24_00017	1	false	527	true	36
aime24_00017	2	false	392	true	696
aime24_00017	3	false	341	true	8712
aime24_00018	0	false	941	true	186
aime24_00018	1	false	3,072	false	null
aime24_00018	2	false	3,072	false	null
aime24_00018	3	false	622	true	312
aime24_00019	0	false	239	true	2
aime24_00019	1	false	514	true	14402
aime24_00019	2	false	1,339	true	2
aime24_00019	3	false	442	true	5
aime24_00020	0	false	3,072	false	null
aime24_00020	1	false	3,072	false	null
aime24_00020	2	false	3,072	false	null
aime24_00020	3	false	3,072	false	null
aime24_00021	0	false	1,328	true	300
aime24_00021	1	false	3,072	false	null
aime24_00021	2	false	903	true	360
aime24_00021	3	false	458	true	300
aime24_00022	0	false	875	true	80
aime24_00022	1	false	1,325	true	80
aime24_00022	2	false	2,232	true	60.8
aime24_00022	3	false	1,582	true	48
aime24_00023	0	false	2,586	true	72
aime24_00023	1	false	1,216	true	206
aime24_00023	2	false	3,072	false	null
aime24_00023	3	false	3,072	false	null
aime24_00024	0	false	901	true	234
aime24_00024	1	false	3,072	false	null
aime24_00024	2	false	1,836	true	1352
aime24_00024	3	false	1,786	true	10

End of preview.

FQ-Select math distillation pool

Data released with the paper Selecting Reasoning Traces by Familiarity and Quality for Small Language Model Distillation (Vu, Nguyen, Nguyen Le; Ton Duc Thang University). Code: https://github.com/JasonNguyen31/QD-RSR-Project (version used for the paper archived at https://doi.org/10.5281/zenodo.24128917). The 8-bit FQ-Select v2 student: https://huggingface.co/VietNguyenHoang/Llama-FQSelect-v2-Qwen2.5-1.5B-Q8_0-GGUF

The release lets others check every number in the paper without calling any API. The teacher endpoints and the LLM judge may change over time, so the solutions and judge scores cannot be regenerated exactly; they are released as they were used.

What is in it

File	Content
`questions.jsonl`	2,000 training questions (GSM8K and MATH train splits, MATH-500 removed). 1,895 of them have at least three correct, judged solutions and form the pool used in the paper.
`candidates.jsonl`	Every correct solution written by the three teachers (DeepSeek-V3, Qwen2.5-VL-72B-Instruct, Llama-3.3-70B-Instruct; three samples per question, temperature 0.8, at most 3,072 tokens): `qid`, `tid`, `teacher`, `sample_idx`, `text`. 15,010 solutions; the 14,927 used in the paper are those of the 1,895 pool questions that the judge could score.
`scores.jsonl`	One row per solution (`tid`). Columns starting with `quality.` are the quality scores (rule half, judge half from Gemini 2.5 Flash-Lite, their min-max normalised values and `qual`). Columns starting with `student_qwen2.5-1.5b.` are the student-side signals computed with the base Qwen2.5-1.5B (`grape` is the familiarity score, mean token log-probability; also RSR, LARK inputs and Local Naturalness where available).
`selections/<rule>.jsonl`	The training set chosen by each selection rule: `qid`, `tid`, three solutions per question, 5,685 rows (5,628 for the rules on the pool without Qwen, which has 1,876 questions). Selections do not depend on the training seed. `selections/<rule>.json` gives the teacher shares and lengths of that set.
`eval/<rule>/seed<N>/<benchmark>.jsonl`	Every evaluation sample of every fine-tuned student: `qid`, `sample` (0 to 3), `correct`, plus token and stop information where recorded. Generated text is not included. Benchmarks: `math500`, `amc23`, `aime24`, `aime25`, `val` (the 200-question validation set), and for some rules `gsm8k`, `bbh`. `summary.json` holds Acc@4, Pass@4 and the decoding settings.
`MANIFEST.json`	md5, size and line count of every file.

Rule names follow the code: correct_only is random selection, grape is GRAPE, fq_select is FQ-Select v1, fq_veto_judge is FQ-Select v2, familiarity_within_teacher is FWT, quality_within_teacher is QWT, teacher_qwen_first and teacher_qwen_last are Prefer Qwen and Avoid Qwen, fit_quality is QD-RSR, *_noqwen are rules on the pool without Qwen, and random_at_grape is random selection within each teacher at GRAPE's teacher shares.

Reproducing a paired difference

Acc@4 of a rule on a question is the mean of correct over its four samples. The paper's paired difference between two rules on one seed is the mean over MATH-500 questions of the difference of their per-question Acc@4, with a 95% interval from 2,000 bootstrap resamples of the questions (src/tools/eval_table.py in the code repository).

Licences and attribution

This dataset combines material under different terms. Users are responsible for complying with each of them.

Questions: GSM8K (MIT) and MATH (MIT).
Solutions written by DeepSeek-V3, accessed through OpenRouter as deepseek/deepseek-chat; recent DeepSeek-V3 releases are under the MIT licence.
Solutions written by Qwen2.5-VL-72B-Instruct, subject to the Qwen License Agreement. Built with Qwen.
Solutions written by Llama-3.3-70B-Instruct. Built with Llama. Llama 3.3 is licensed under the Llama 3.3 Community License, Copyright © Meta Platforms, Inc. All Rights Reserved. Any model trained on these solutions and then distributed must follow that licence, including the requirement that its name begin with "Llama".
Judge scores were produced by Gemini 2.5 Flash-Lite through OpenRouter.
Scores, selections, evaluation results and this card were produced by the authors and are released under CC BY 4.0.

Citation

The paper is under review; citation details will be added here once it is published.

Downloads last month: 15

Total file size:

79.2 MB

Models trained or fine-tuned on VietNguyenHoang/fq-select-math-distillation

VietNguyenHoang/Llama-FQSelect-v2-Qwen2.5-1.5B-Q8_0-GGUF

Text Generation • 2B • Updated about 8 hours ago