xent-labs/GooseReason-0.7M

Xent GooseReason 0.7M

This is a reproducibly shuffled and benchmark-decontaminated derivative of `nvidia/Nemotron-Research-GooseReason-0.7M`.

Splits

Each source subset has a training split plus validation and test splits containing 500 rows each. Holdouts are stratified by num_choices with largest-remainder allocation, then every split is deterministically shuffled with seed 42.

Published splits are math_train, math_val, math_test, code_train, code_val, code_test, stem_train, stem_val, and stem_test.

Rows preserve question, options, and answer and add a stable row id, a source_question_id, and num_choices. All rows derived from the same normalized source question remain together. Validation and test are selected from singleton source-question groups, while repeated-question groups remain in training.

Only rows containing exactly one [MASK] after the **Solution:** marker are retained. Rows with no solution mask or multiple solution masks are recorded in removed_invalid.jsonl and excluded before the stratified split is created.

Contamination policy

Every source question was extracted from the **Question:** section and compared against all published splits of:

• —openai/gsm8k (train, test)
• —hendrydong/hendrycks_math (train, test, math500)

Only exact equality after conservative presentation normalization is used. The normalizer applies Unicode NFKC, removes whitespace and punctuation through tokenization, drops presentation-only LaTeX commands, and folds ordinary prose words to lowercase. Numbers, operators, single-letter variable case, uppercase symbolic names, and content-bearing LaTeX commands are preserved. There is no fuzzy retrieval or similarity-based removal.

contamination_report.json contains aggregate results and removed_contamination.jsonl records every removal and its matched benchmark rows. removed_invalid.jsonl records malformed source rows removed before split assignment.

Reproduction

The preparation script lives in the associated project repository and uses source row order plus seed 42. Source dataset revisions and library versions are recorded in contamination_report.json.

License

The source dataset is licensed CC BY-NC 4.0. This derivative retains that license.