Datasets:

BASIC-MATH
/

math-rl-qwen35-k8

Tags:

Dataset Viewer

Split (1)

train

The dataset viewer is not available for this split.

Cannot load the dataset split (in streaming mode) to extract the first rows.

Error code:   StreamingRowsError
Exception:    CastError
Message:      Couldn't cast
uid: string
question: string
gt: string
rubrics: list<item: struct<criterion: string, points: double, tags: struct<judge: bool, inject: bool, scaffol (... 84 chars omitted)
  child 0, item: struct<criterion: string, points: double, tags: struct<judge: bool, inject: bool, scaffold_rank: int (... 72 chars omitted)
      child 0, criterion: string
      child 1, points: double
      child 2, tags: struct<judge: bool, inject: bool, scaffold_rank: int64, scaffold_kind: string, verify: string, rewar (... 23 chars omitted)
          child 0, judge: bool
          child 1, inject: bool
          child 2, scaffold_rank: int64
          child 3, scaffold_kind: string
          child 4, verify: string
          child 5, reward: bool
          child 6, value: string
lang: string
split: string
scores: struct<p_hat: double, k: int64, n_trunc: int64, model: string, temp: double, max_tokens: int64, hits (... 211 chars omitted)
  child 0, p_hat: double
  child 1, k: int64
  child 2, n_trunc: int64
  child 3, model: string
  child 4, temp: double
  child 5, max_tokens: int64
  child 6, hits: list<item: int64>
      child 0, item: int64
  child 7, toks: list<item: int64>
      child 0, item: int64
  child 8, gen_cap: list<item: int64>
      child 0, item: int64
  child 9, by_cap: struct<32768: struct<p_hat: double, n_trunc: int64, class: string>, 65536: struct<p_hat: double, n_t (... 28 chars omitted)
      child 0, 32768: struct<p_hat: double, n_trunc: int64, class: string>
          child 0, p_hat: double
          child 1, n_trunc: int64
          child 2, class: string
      child 1, 65536: struct<p_hat: double, n_trunc: int64, class: string>
          child 0, p_hat: double
          child 1, n_trunc: int64
          child 2, class: string
val: int64
train: int64
learnable_by_cap: struct<32768: int64, 65536: int64>
  child 0, 32768: int64
  child 1, 65536: int64
rows: int64
repo: string
case: string
files: struct<math_rl_qwen35_k8.jsonl: struct<sha256_12: string, bytes: int64>, rubrics/rubric.jsonl: struc (... 87 chars omitted)
  child 0, math_rl_qwen35_k8.jsonl: struct<sha256_12: string, bytes: int64>
      child 0, sha256_12: string
      child 1, bytes: int64
  child 1, rubrics/rubric.jsonl: struct<sha256_12: string, bytes: int64>
      child 0, sha256_12: string
      child 1, bytes: int64
  child 2, reward.py: struct<sha256_12: string, bytes: int64>
      child 0, sha256_12: string
      child 1, bytes: int64
rubric_rows: int64
to
{'case': Value('string'), 'repo': Value('string'), 'rows': Value('int64'), 'train': Value('int64'), 'val': Value('int64'), 'rubric_rows': Value('int64'), 'learnable_by_cap': {'32768': Value('int64'), '65536': Value('int64')}, 'files': {'math_rl_qwen35_k8.jsonl': {'sha256_12': Value('string'), 'bytes': Value('int64')}, 'rubrics/rubric.jsonl': {'sha256_12': Value('string'), 'bytes': Value('int64')}, 'reward.py': {'sha256_12': Value('string'), 'bytes': Value('int64')}}}
because column names don't match
Traceback:    Traceback (most recent call last):
                File "/src/services/worker/src/worker/utils.py", line 147, in get_rows_or_raise
                  return get_rows(
                      dataset=dataset,
                  ...<4 lines>...
                      column_names=column_names,
                  )
                File "/src/libs/libcommon/src/libcommon/utils.py", line 272, in decorator
                  return func(*args, **kwargs)
                File "/src/services/worker/src/worker/utils.py", line 127, in get_rows
                  rows_plus_one = list(itertools.islice(safe_iter(ds, dataset=dataset), rows_max_number + 1))
                File "/src/services/worker/src/worker/utils.py", line 483, in safe_iter
                  yield from ds.decode(False) if ds.features else ds
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2840, in __iter__
                  for key, example in ex_iterable:
                                      ^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2373, in __iter__
                  for key, pa_table in self._iter_arrow():
                                       ~~~~~~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2398, in _iter_arrow
                  for key, pa_table in self.ex_iterable._iter_arrow():
                                       ~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 536, in _iter_arrow
                  for key, pa_table in iterator:
                                       ^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 419, in _iter_arrow
                  for key, pa_table in self.generate_tables_fn(**gen_kwags):
                                       ~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^
                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
              uid: string
              question: string
              gt: string
              rubrics: list<item: struct<criterion: string, points: double, tags: struct<judge: bool, inject: bool, scaffol (... 84 chars omitted)
                child 0, item: struct<criterion: string, points: double, tags: struct<judge: bool, inject: bool, scaffold_rank: int (... 72 chars omitted)
                    child 0, criterion: string
                    child 1, points: double
                    child 2, tags: struct<judge: bool, inject: bool, scaffold_rank: int64, scaffold_kind: string, verify: string, rewar (... 23 chars omitted)
                        child 0, judge: bool
                        child 1, inject: bool
                        child 2, scaffold_rank: int64
                        child 3, scaffold_kind: string
                        child 4, verify: string
                        child 5, reward: bool
                        child 6, value: string
              lang: string
              split: string
              scores: struct<p_hat: double, k: int64, n_trunc: int64, model: string, temp: double, max_tokens: int64, hits (... 211 chars omitted)
                child 0, p_hat: double
                child 1, k: int64
                child 2, n_trunc: int64
                child 3, model: string
                child 4, temp: double
                child 5, max_tokens: int64
                child 6, hits: list<item: int64>
                    child 0, item: int64
                child 7, toks: list<item: int64>
                    child 0, item: int64
                child 8, gen_cap: list<item: int64>
                    child 0, item: int64
                child 9, by_cap: struct<32768: struct<p_hat: double, n_trunc: int64, class: string>, 65536: struct<p_hat: double, n_t (... 28 chars omitted)
                    child 0, 32768: struct<p_hat: double, n_trunc: int64, class: string>
                        child 0, p_hat: double
                        child 1, n_trunc: int64
                        child 2, class: string
                    child 1, 65536: struct<p_hat: double, n_trunc: int64, class: string>
                        child 0, p_hat: double
                        child 1, n_trunc: int64
                        child 2, class: string
              val: int64
              train: int64
              learnable_by_cap: struct<32768: int64, 65536: int64>
                child 0, 32768: int64
                child 1, 65536: int64
              rows: int64
              repo: string
              case: string
              files: struct<math_rl_qwen35_k8.jsonl: struct<sha256_12: string, bytes: int64>, rubrics/rubric.jsonl: struc (... 87 chars omitted)
                child 0, math_rl_qwen35_k8.jsonl: struct<sha256_12: string, bytes: int64>
                    child 0, sha256_12: string
                    child 1, bytes: int64
                child 1, rubrics/rubric.jsonl: struct<sha256_12: string, bytes: int64>
                    child 0, sha256_12: string
                    child 1, bytes: int64
                child 2, reward.py: struct<sha256_12: string, bytes: int64>
                    child 0, sha256_12: string
                    child 1, bytes: int64
              rubric_rows: int64
              to
              {'case': Value('string'), 'repo': Value('string'), 'rows': Value('int64'), 'train': Value('int64'), 'val': Value('int64'), 'rubric_rows': Value('int64'), 'learnable_by_cap': {'32768': Value('int64'), '65536': Value('int64')}, 'files': {'math_rl_qwen35_k8.jsonl': {'sha256_12': Value('string'), 'bytes': Value('int64')}, 'rubrics/rubric.jsonl': {'sha256_12': Value('string'), 'bytes': Value('int64')}, 'reward.py': {'sha256_12': Value('string'), 'bytes': Value('int64')}}}
              because column names don't match

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.

math-rl-qwen35-k8 — 데이터 · 루브릭 · 리워드 한 묶음

대상 모델 Qwen3.5-35B-A3B(thinking)로 후보 풀 63,528문제를 선별해, 그 모델이 풀다 말다 하는 문제만 남긴 GRPO 학습 케이스다. 구성은 BASIC-MATH/t01-q35-bin-v1 과 같다.

응답 상한을 케이스에 박지 않았다. 표본별 생성 토큰 수를 행에 담고 상한별 판정을 함께 넣었으니, 32K 로 학습하든 64K 로 학습하든 같은 키로 걸러 쓴다.

항목	값
데이터	`math_rl_qwen35_k8.jsonl` 7,525행 (train 7,480 / val 45)
선별	롤아웃 K=4~8, temp 0.7 · top-p 0.95. 32K 패스 + 64K 재측정 패스 합산
루브릭	`rubrics/rubric.jsonl` 7,046행 — 중간값 기준(`contains_value`, reward, inject=false, rank 9) + 교사 힌트 3종(`strategy`/`subgoal`/`step_hint`, inject=true, rank 1·1·0)
루브릭 커버리지	값 기준 6,283행 · 힌트 4,823행 · 둘 다 없음 479행
리워드	`reward.py` — 이진 정답(정답 1 · 그 외 0 · 사고 미완료 0) + 잘림 loss-masking(`grpo_mask_trunc`)

상한별로 쓰는 법 ★

행마다 scores 에 이런 키가 들어 있다.

"hits":    [0, 0, 0, 0, 0, 1, 1, 0],
"toks":    [32768, 32768, 32768, 32768, 52533, 59054, 29006, 65536],
"gen_cap": [32768, 32768, 32768, 32768, 65536, 65536, 65536, 65536],
"by_cap": {
  "32768": {"p_hat": 0.125, "n_trunc": 7, "class": "learnable_no_wrong"},
  "65536": {"p_hat": 0.250, "n_trunc": 5, "class": "learnable"}
}

toks 는 표본별 생성 토큰 수, gen_cap 은 그 표본을 만든 패스의 생성 상한이다.
학습 응답 32K 로 돌린다면 scores.by_cap["32768"].class == "learnable" 인 행만 쓴다 → 6,378행
64K 로 돌린다면 scores.by_cap["65536"].class == "learnable" 인 행을 쓴다 → 7,480행

위 예시는 52,533·59,054 토큰짜리 정답이 있어 32K 로 읽으면 보류, 64K 로 읽으면 채택이 되는 문제다. 상한을 바꾸면 채택 집합이 달라진다 — 잘린 표본의 정답을 인정하지 않기 때문이다. 학습 상한과 거르는 키를 반드시 맞춰라.

선별 판정 (screened 63,528)

부류	상한 32,768	상한 65,536
learnable	6,378	7,480
learnable_no_wrong	2,439	2,198
unresolved (전부 잘림)	6,972	1,224
dead	42,159	47,046
solved	5,580	5,580

64K 재측정이 도는 동안 30분마다 갱신된다. 진행될수록 두 열의 격차가 벌어진다.

붙이는 법

reward_model.ground_truth = gt        (answer 아님)
extra_info.index          = uid
extra_info.rubrics        = rubrics   (행에 들어 있는 list 그대로)
custom_reward_function.path = reward.py · name = compute_score_dict
calculate_log_probs=False · static_pack 지정 금지

rubrics 는 데이터 행 안에도, rubrics/rubric.jsonl 에도 같은 내용이 들어 있다(빌드 때 같은 원본).

주의

루브릭이 비어 있는 479행은 참조 풀이를 구하지 못한 것이다. 이진 리워드 학습에는 영향이 없다. 대상 모델의 정답 풀이로 생성해 채우는 중이다.
p_hat·n_trunc·max_tokens 최상위 키는 상한 32,768 기준(하위호환)이다. 판정은 by_cap 을 봐라.

Downloads last month: 32

Total file size:

26.1 MB

math-rl-qwen35-k8 — 데이터 · 루브릭 · 리워드 한 묶음

상한별로 쓰는 법 ★

선별 판정 (screened 63,528)

붙이는 법

관련 리포

주의