Example for training on SWE (agentic software-engineering) tasks?

[dipta007]

Is there an example/recipe for grpo training on multi-turn SWE-style tasks (SWE-bench / SWE-Gym)?

Looking for:

• Multi-turn agentic rollouts with tool calls (shell, file edits, tests)
• How environment/reward is wired (e.g. tests-passing as reward)
• Setup for long trajectories / large context
• Setup for how docker/podman is handled efficiently during rollout

If nothing SWE-specific exists, a pointer to the closest multi-turn example to adapt would help. Happy to contribute one back.

Thanks!

[chenyushuo]

Hi, thanks for the question.

At the moment, we do not have an official GRPO recipe specifically for multi-turn SWE-style tasks (such as SWE-bench or SWE-Gym) in this repository.

The closest reference is our multi-turn agent training example based on QwenPaw which is a OpenClaw-like AI assistant, and this example also uses GRPO:
https://github.com/chenyushuo/Trinity-RFT/tree/dev/copaw_rl/examples/copaw_rl

More specifically:

• Workflows used for Trinity training are under:
  https://github.com/chenyushuo/Trinity-RFT/tree/dev/copaw_rl/examples/copaw_rl/workflows
• Utilities for running QwenPaw inside E2B Sandbox, trajectory scoring, and trajectory export are under:
  https://github.com/chenyushuo/Trinity-RFT/tree/dev/copaw_rl/examples/copaw_rl/utils

In this setup, the workflow runs QwenPaw through E2B Sandbox, and the utils layer handles rollout execution, reward/scoring, and trajectory processing inside E2B Sandbox. While it is not SWE-specific out of the box, it should be a practical starting point for adapting to SWE-style multi-turn training (tool use, environment wiring, and long-horizon trajectories).

If you are open to contributing an SWE-focused recipe, that would be very welcome.