Skip to content

Link the TRL example and the multi-harness article from harbor_env #1293

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
sergiopaniego wants to merge 2 commits into
base: main
Choose a base branch
from
Open

Link the TRL example and the multi-harness article from harbor_env #1293

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
fdfc05d
Link the TRL example and the multi-harness article from harbor_env
sergiopaniego Oct 1, 2026
06c4af3
Link the FineEnvs multi-harness tutorial from harbor_env
sergiopaniego Oct 1, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
17 changes: 17 additions & 0 deletions docs/source/environments/harbor.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -259,6 +259,23 @@ with HarborEnv(base_url="http://localhost:8000") as env:
`harness` and `sandbox` are per call, so consecutive rollouts against the same server can use
different agents and different backends.

## Training with TRL

TRL trains a policy on these rollouts with `AsyncGRPOTrainer` (on TRL `main` for now): a
`HarnessRolloutWorker` builds a `HarborSessionFactory` with its sampling policy, opens one session
per rollout and trains on each session's validated `TrainingTrace`.

- [`examples/async_grpo_harbor`](https://github.com/huggingface/trl/tree/main/examples/async_grpo_harbor):
a complete training script, with local setup and a Hugging Face Jobs launcher.
- [TRL's OpenEnv guide](https://huggingface.co/docs/trl/openenv): how the worker, the reward and the
capture fit together.
- [The ultimate guide to multi-harness RL](https://huggingface.co/spaces/AdithyaSK/multi-harness-rl):
one policy trained across OpenCode, Claude Code, Codex and Mini-SWE-Agent through this
environment. Its runnable code is the
[FineEnvs multi-harness tutorial](https://github.com/adithya-s-k/FineEnvs/tree/main/05-multi-harness-rl),
and its [models, datasets and environments](https://huggingface.co/collections/FineEnvs/smoldataenv-multi-harness-rl)
are in one collection.

## The web UI

`serve` (and a Space made with `push`) serves a UI at `/web`, in four tabs.
Expand Down
17 changes: 17 additions & 0 deletions envs/harbor_env/README.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -267,6 +267,23 @@ with HarborEnv(base_url="http://localhost:8000") as env:
`harness` and `sandbox` are per call, so consecutive rollouts against the same server can use
different agents and different backends.

## Training with TRL

TRL trains a policy on these rollouts with `AsyncGRPOTrainer` (on TRL `main` for now): a
`HarnessRolloutWorker` builds a `HarborSessionFactory` with its sampling policy, opens one session
per rollout and trains on each session's validated `TrainingTrace`.

- [`examples/async_grpo_harbor`](https://github.com/huggingface/trl/tree/main/examples/async_grpo_harbor):
a complete training script, with local setup and a Hugging Face Jobs launcher.
- [TRL's OpenEnv guide](https://huggingface.co/docs/trl/openenv): how the worker, the reward and the
capture fit together.
- [The ultimate guide to multi-harness RL](https://huggingface.co/spaces/AdithyaSK/multi-harness-rl):
one policy trained across OpenCode, Claude Code, Codex and Mini-SWE-Agent through this
environment. Its runnable code is the
[FineEnvs multi-harness tutorial](https://github.com/adithya-s-k/FineEnvs/tree/main/05-multi-harness-rl),
and its [models, datasets and environments](https://huggingface.co/collections/FineEnvs/smoldataenv-multi-harness-rl)
are in one collection.

## The web UI

`serve` (and a Space made with `push`) serves a UI at `/web`, in four tabs.
Expand Down
Loading