Dataset Viewer
Duplicate
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
num_problems: int64
num_samples_per_problem: int64
pass@1: struct<mean: double, lo: double, hi: double, n: int64>
  child 0, mean: double
  child 1, lo: double
  child 2, hi: double
  child 3, n: int64
mean_sample_accuracy: struct<mean: double, lo: double, hi: double, n: int64>
  child 0, mean: double
  child 1, lo: double
  child 2, hi: double
  child 3, n: int64
truncation_rate: double
extraction_failure_rate: double
tag_format_rate: double
extraction_sources: struct<tag: int64, final_phrase: int64, last_number: int64, boxed: int64>
  child 0, tag: int64
  child 1, final_phrase: int64
  child 2, last_number: int64
  child 3, boxed: int64
tokens: struct<mean: double, median: double, p90: double>
  child 0, mean: double
  child 1, median: double
  child 2, p90: double
tokens_correct: struct<mean: double, median: double, p90: double>
  child 0, mean: double
  child 1, median: double
  child 2, p90: double
tokens_incorrect: struct<mean: double, median: double, p90: double>
  child 0, mean: double
  child 1, median: double
  child 2, p90: double
pass@k: struct<1: struct<mean: double, lo: double, hi: double, n: int64>, 2: struct<mean: double, lo: double (... 142 chars omitted)
  child 0, 1: struct<mean: double, lo: double, hi: double, n: int64>
      child 0, mean: double
      child 1, lo: double
      child 2, hi: double
      child 3, n: int64
  child 1, 2: struct<mean: double, lo: double, hi: double, n: int64>
      child 0, mean: double
      child 1, lo: double
      child 
...
2: double
  child 2, 4: double
  child 3, 8: double
eval: string
template: string
model: string
run_tag: string
sampling: struct<n: int64, temperature: double, top_p: double, top_k: int64, max_tokens: int64, seed: int64, s (... 24 chars omitted)
  child 0, n: int64
  child 1, temperature: double
  child 2, top_p: double
  child 3, top_k: int64
  child 4, max_tokens: int64
  child 5, seed: int64
  child 6, stop: list<item: string>
      child 0, item: string
manifest: struct<name: string, source: string, revision: null, answer_type: string, num_problems: int64, conte (... 16 chars omitted)
  child 0, name: string
  child 1, source: string
  child 2, revision: null
  child 3, answer_type: string
  child 4, num_problems: int64
  child 5, content_hash: string
wall_clock_s: double
backend: struct<backend: string, model: string, revision: null, vllm: string, batch_invariant: bool>
  child 0, backend: string
  child 1, model: string
  child 2, revision: null
  child 3, vllm: string
  child 4, batch_invariant: bool
context_budget: struct<max_model_len: int64, max_tokens: int64, prompt_tokens_min: int64, prompt_tokens_max: int64,  (... 90 chars omitted)
  child 0, max_model_len: int64
  child 1, max_tokens: int64
  child 2, prompt_tokens_min: int64
  child 3, prompt_tokens_max: int64
  child 4, effective_max_tokens_min: int64
  child 5, effective_max_tokens_max: int64
  child 6, capped_by_context: bool
source: string
answer_type: string
revision: null
name: string
content_hash: string
to
{'name': Value('string'), 'source': Value('string'), 'revision': Value('null'), 'answer_type': Value('string'), 'num_problems': Value('int64'), 'content_hash': Value('string')}
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 2951, in __iter__
                  for key, example in ex_iterable:
                                      ^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2461, in __iter__
                  for key, pa_table in self._iter_arrow():
                                       ~~~~~~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2486, 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 547, in _iter_arrow
                  for key, pa_table in iterator:
                                       ^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 430, 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
              num_problems: int64
              num_samples_per_problem: int64
              pass@1: struct<mean: double, lo: double, hi: double, n: int64>
                child 0, mean: double
                child 1, lo: double
                child 2, hi: double
                child 3, n: int64
              mean_sample_accuracy: struct<mean: double, lo: double, hi: double, n: int64>
                child 0, mean: double
                child 1, lo: double
                child 2, hi: double
                child 3, n: int64
              truncation_rate: double
              extraction_failure_rate: double
              tag_format_rate: double
              extraction_sources: struct<tag: int64, final_phrase: int64, last_number: int64, boxed: int64>
                child 0, tag: int64
                child 1, final_phrase: int64
                child 2, last_number: int64
                child 3, boxed: int64
              tokens: struct<mean: double, median: double, p90: double>
                child 0, mean: double
                child 1, median: double
                child 2, p90: double
              tokens_correct: struct<mean: double, median: double, p90: double>
                child 0, mean: double
                child 1, median: double
                child 2, p90: double
              tokens_incorrect: struct<mean: double, median: double, p90: double>
                child 0, mean: double
                child 1, median: double
                child 2, p90: double
              pass@k: struct<1: struct<mean: double, lo: double, hi: double, n: int64>, 2: struct<mean: double, lo: double (... 142 chars omitted)
                child 0, 1: struct<mean: double, lo: double, hi: double, n: int64>
                    child 0, mean: double
                    child 1, lo: double
                    child 2, hi: double
                    child 3, n: int64
                child 1, 2: struct<mean: double, lo: double, hi: double, n: int64>
                    child 0, mean: double
                    child 1, lo: double
                    child 
              ...
              2: double
                child 2, 4: double
                child 3, 8: double
              eval: string
              template: string
              model: string
              run_tag: string
              sampling: struct<n: int64, temperature: double, top_p: double, top_k: int64, max_tokens: int64, seed: int64, s (... 24 chars omitted)
                child 0, n: int64
                child 1, temperature: double
                child 2, top_p: double
                child 3, top_k: int64
                child 4, max_tokens: int64
                child 5, seed: int64
                child 6, stop: list<item: string>
                    child 0, item: string
              manifest: struct<name: string, source: string, revision: null, answer_type: string, num_problems: int64, conte (... 16 chars omitted)
                child 0, name: string
                child 1, source: string
                child 2, revision: null
                child 3, answer_type: string
                child 4, num_problems: int64
                child 5, content_hash: string
              wall_clock_s: double
              backend: struct<backend: string, model: string, revision: null, vllm: string, batch_invariant: bool>
                child 0, backend: string
                child 1, model: string
                child 2, revision: null
                child 3, vllm: string
                child 4, batch_invariant: bool
              context_budget: struct<max_model_len: int64, max_tokens: int64, prompt_tokens_min: int64, prompt_tokens_max: int64,  (... 90 chars omitted)
                child 0, max_model_len: int64
                child 1, max_tokens: int64
                child 2, prompt_tokens_min: int64
                child 3, prompt_tokens_max: int64
                child 4, effective_max_tokens_min: int64
                child 5, effective_max_tokens_max: int64
                child 6, capped_by_context: bool
              source: string
              answer_type: string
              revision: null
              name: string
              content_hash: string
              to
              {'name': Value('string'), 'source': Value('string'), 'revision': Value('null'), 'answer_type': Value('string'), 'num_problems': Value('int64'), 'content_hash': Value('string')}
              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.

reasoning-rlvr: per-sample outputs

Every completion behind the numbers of the reasoning-rlvr project: SFT, rejection-sampling fine-tuning (RFT) and GRPO with a verifiable reward on Qwen3-1.7B-Base, evaluated on GSM8K, MATH-500 and Countdown, plus the test-time-compute experiments that followed (forced stopping, "Wait", an outcome reward model). The checkpoints are Raviteja546/reasoning-rlvr-sft, -rft and -grpo.

Sample-set folders

Each top-level folder is one sampled run, <run>/<eval>/:

  • samples.jsonl: one record per completion: problem_id, sample_idx, text (the completion), num_tokens, finish_reason (stop or length), extracted (the parsed answer or null), extraction_source (tag, boxed, final_phrase, last_number or none), correct, format_ok, gold, meta. A completion whose answer no rule extracts is scored wrong.
  • metrics.json, manifest.json (problem ids and their hash), prompt_example.txt; run_info.json at the run level (model, commit, sampling settings).
  • scores.jsonl (in the three *_n32_think_tag_t06 folders, GSM8K and MATH-500, added 24 Oct 2026): the outcome reward model's score (logit) and prob per completion, with correct for convenience. The reward model is a LoRA adapter (r 16, alpha 32) and a scalar head on Qwen3-0.6B-Base, trained on 20,000 labelled SFT samples; held-out accuracy 89.4 %, AUROC 0.944. Its weights are in the laptop archive below.

*_n32_think_tag_t06: T = 0.6, top-p 0.95, 32 samples per problem on pinned 300-problem subsets. *_n64_*: 64 samples per problem. sft_1.7b_rft_n8: the solve-rate pass that built the RFT set (and the reward model's training data).

Archives (archives/)

Each archive extracts to an outputs/ tree; the .sha256 files hold the hashes.

archive what it holds
h100_phase2_archive.tar.gz the Phase 2 (SFT) H100 session
h100_session_archive.tar.gz the Phase 2-3 H100 session, 18-19 Oct 2026: every metrics file, greedy per-sample file, report and RFT manifest
h100_p4_session1_archive.tar.gz, _addendum GRPO session 1, 20-21 Oct 2026
h100_p4_session2_archive.tar.gz, _addendum, h100_p4_session2_addendum2.tar.gz, _verify.txt GRPO session 2, 21 Oct 2026
laptop_archive_2026-10-24.tar.gz everything run on the laptop (RTX 4060): the base-model greedy runs and probes of Phase 1, the forced-stop runs (*_force{768..3072}_think_tag), the "Wait" runs, the same-machine unforced controls (*_laptop_think_tag), the outcome reward model (outputs/train/orm_0p6b: adapter, head, metrics, training log) and its training data (outputs/data/orm), and the run logs

Note on pairing

Greedy decoding with the project's batch-invariant harness is bit-identical on one machine, not across GPUs: the same checkpoint decoded on an H100 and on the laptop flips 10-23 % of problems. Compare a laptop run with a laptop run and an H100 run with an H100 run.

Downloads last month
192