AXCXEPT/

tmax-balanced-open-instruct-v1

This dataset merges TMax-compatible verifier-backed RL task datasets into one Open-Instruct/SWERL Vanillux sandbox dataset. Schema:

messages ground_truth dataset env_config source

Rows: 38,577 Balance profile: balanced.

SWE-Smith cap: 10,000 rows Selection seed: 20260624

Sources:

allenai/tmax-15k-open-instruct: 14,601 rows allenai/open-instruct-swe-smith: 10,000 rows allenai/open-instruct-terminal-traj: 5,648 rows… See the full description on the dataset page: https://huggingface.co/datasets/AXCXEPT/tmax-balanced-open-instruct-v1.