README.md

metadata

license: odc-by
language:
  - en
tags:
  - tmax
  - open-instruct
  - reinforcement-learning
  - terminal-agents
  - swerl_vanillux_sandbox

TMax Mixed Open-Instruct RL Dataset 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
• allenai/open-instruct-termigen: 3,556 rows
• allenai/open-instruct-cli-gym: 1,552 rows
• allenai/open-instruct-endless-terminals: 2,492 rows
• allenai/open-instruct-openthoughts: 728 rows

The included task-data.tar.gz contains the corresponding task directories. Task ids are prefixed per source to avoid collisions.