schneiderkamplab/dfm11-mathagentic-tinygsm-python

dfm11-mathagentic-tinygsm-python

English arithmetic word problems converted into native Python tool-call trajectories with precomputed tool responses and boxed final answers.

Contents

• —Rows: 367,749
• —Shards: 4
• —Format: deterministic gzip JSON Lines in data/train-*.jsonl.gz
• —Schema: tools, four-message native tool trajectory, execution metadata, stable source ID, source revision, and admission status
• —Intended repository: schneiderkamplab/dfm11-mathagentic-tinygsm-python

Each row contains a user problem, an assistant tool call, a precomputed tool response, and a terminal assistant response containing exactly one boxed final answer. It is directly consumable by Mathagentic/HRM-Text's Gemma 4 native tool-call rendering path; no executor is used during training.

Upstream datasets

• —TinyGSM/TinyGSM

Selection and verification

Deterministic 1-in-20 source sample; restricted-Python validation and execution; independent Gemma 4 26B A4B semantic audit; accepted verdicts only.

The semantic audit covered 500,000 candidates and accepted 367,749 (73.55%). Rejected rows are not packaged.

The package contains source trajectories, not tokenized arrays or sampled epoch indices. Packaging does not itself admit the dataset into a DFM11 data mix. The approved DFM11 sampling policy is to deterministically cap to 50,000 distinct rows per epoch, with repeat 1. This policy does not remove rows from the published source package.

License and citation review

The upstream TinyGSM dataset declares the MIT license. Preserve its attribution.

Validate

python validate_dataset.py