Scaling RL for language models depends on diverse, verifiable environments that produce reliable reward signals. MiMo-V2.6-RL-oss supplies a curated suite of agentic RL environments spanning software engineering, vulnerability reproduction, knowledge-intensive tasks, web/visual dev and symbolic music, packaged as dataset configs plus runtime artifacts so teams can reproduce and extend mixed-task RL runs without rebuilding basic harnesses.
What Sets It Apart
- Multi-domain, verifier-driven design: each domain pairs tasks with a concrete verifier type (executable tests for code, rule checks for cyber/music, rubric-based judging for knowledge work, visual grading for web dev) so rewards are rooted in automated, domain-appropriate checks.
- Ready-to-run artifacts: datasets published in Parquet with named configs (code, cyber, general, webdev, music), plus Docker images and pointers to the verl training code to simplify environment orchestration and high-concurrency rollouts.
- Built for mixed-task, agentic RL: organized to support mixed-harness training where tasks and harnesses interleave in the same batch, enabling capability transfer across domains and grader-driven groupwise reward mechanisms.
- Open license and standard formats: Apache-2.0 license and Parquet/standard toolchain make the assets easy to inspect, reuse, and integrate into existing RL pipelines.
Who It's For & Tradeoffs
Great fit if you are a researcher or engineering team building or evaluating agentic RL for LLMs and need reproducible, verifiable task environments across coding, security, knowledge-work, web visual design, and symbolic music. It lowers engineering overhead by providing verifiers, Docker images and dataset configs. Look elsewhere if you need a hosted interactive environment platform, GUIs for human-in-the-loop grading, or extremely large-scale commercial environment catalogs not included here—the package focuses on verifiable, reproducible RL harnesses rather than human-annotation services or turnkey cloud hosting.
How It's Structured
Datasets are provided as Parquet files under named configs (code.parquet, cyber.parquet, general/train.parquet, webdev.parquet, music.parquet). The dataset card links to Docker images and the verl training repo for running rollouts, collecting trajectories, and integrating verifier-driven reward evaluation.
