Repo2RLEnv — verifiable RL environments from any repository

Summary

Generates Harbor-format coding, terminal, and reasoning tasks — each with instruction, environment, private verifier, and reference solution — from any Git repository. Tasksmith reads pull requests, builds repositories, writes tasks and verifiers, and self-repairs until controls pass. The site reports 23 pipelines and 21 published Hugging Face datasets.

Keywords

RL environments · verifiable tasks · code · Tasksmith · Hugging Face

Topics

RL environments, verifiable tasks, code, Tasksmith, Hugging Face

Research notes

• Discovery: Shared in #random-papers as a project-site link.
• Method: Tasksmith pipeline: PR mining -> repo build -> task/verifier authoring -> self-repair against controls -> publish as Harbor-format verifiable environments.
• The 21 published HF datasets are named on the site; a datasets-batch subagent could enumerate them for the Datasets sheet.