---
pretty_name: Clean Math SFT Problem Pool
configs:
- config_name: default
  data_files:
  - split: train
    path: pool.jsonl
---
# Clean math SFT problem pool

Inference-free preparation of MATH training, NuminaMath-1.5, and original s1K.
Only questions, extracted gold answers, and provenance are retained; no generated traces.
The `metadata` column contains lossless JSON text for heterogeneous source metadata.
Source licenses remain applicable; see provenance.json for upstream metadata and revisions.
The repository's unchanged AIME 2024/2025/2026, AMC 2023, MATH-500 evaluation
references and its content-checked DAPO RL set are excluded using distinct n-gram coverage.

`summary.json` contains stage counts; `removed.jsonl` records every removal;
`conflicts.jsonl` quarantines conflicting duplicate clusters; `merged_clusters.jsonl`
records merges. The two `*_near_misses.jsonl` files list manual-review candidates
in descending coverage order. Exact number-swapped templates can appear below the
usual near-miss band and are explicitly marked. Eval near misses are recorded at
the eval stage, so a candidate can subsequently be removed for RL overlap.

Filtering uses conservative rules, not a semantic classifier; inspect the audit logs.
Symbolic parsing does not establish that an upstream gold answer is correct.
Lexical normalization removes punctuation, and is not symbolic canonicalization.
Conflicting near-duplicate clusters are dropped in full as a conservative policy.

See `provenance.json` for the complete config, pinned inputs, code commit and
bundled-file hashes. This artifact performs no model inference or training.