task_categories:
- text-generation
Shared math feedback candidate inputs
This archive contains 383,537 unscored generated-answer records and 184,538 online question rows covering the same 184,502 normalized question IDs. It is a candidate input corpus, not a newly judged feedback dataset or an evaluated model.
The pool intersects pre-existing offline and online question IDs and excludes exact normalized matches to five math test sets. Six overlapping questions and 129 offline-only questions were excluded; one empty generated answer was omitted. Duplicate-row weights, source order and original strings are retained. Historical positive conditions were not reused.
Each offline record contains question, model_answer and origin metadata (source line, branch and SHA256). Online records contain question and source row identity. Historical pos/neg branch names are provenance labels, not newly validated correctness judgments. Original generator and selection history remain unverified. Exact overlap exclusion does not establish semantic or pretraining decontamination.
manifest.json records every shard hash, line count and the SHA256 of each concatenated original JSONL file. Shards are in numeric order. Concatenating their gzip-decoded bytes recovers the corresponding JSONL exactly. The shared-pool protocol fingerprint is included for auditability.
No API credentials, model weights, critic judgments, training budget or performance claims are included.