diff --git a/README.md b/README.md index 850f162..458c76c 100644 --- a/README.md +++ b/README.md @@ -39,3 +39,4 @@ harbor run -p harbor_cookbook/recipes/ -a claude-code -m anthropic/claude- |:--|:--| | [gepa](harbor_cookbook/gepa/) | Agent harness optimization for MedAgentBench using Harbor+GEPA. | | [tinker‑rl](harbor_cookbook/tinker_rl/) | RL training on Harbor tasks using the Tinker SDK. | +| [sky‑rl](harbor_cookbook/sky_rl/) | RL training on Harbor tasks using SkyRL. | diff --git a/harbor_cookbook/sky_rl/README.md b/harbor_cookbook/sky_rl/README.md new file mode 100644 index 0000000..7d04ca2 --- /dev/null +++ b/harbor_cookbook/sky_rl/README.md @@ -0,0 +1,31 @@ +# sky_rl + +RL training on Harbor tasks using [SkyRL](https://github.com/NovaSky-AI/SkyRL). + +This directory is a pointer. The full integration lives in the SkyRL repo: + +**[SkyRL/examples/train_integrations/harbor](https://github.com/NovaSky-AI/SkyRL/tree/main/examples/train_integrations/harbor)** + +## Quick start + +Set credentials: + +```bash +export WANDB_API_KEY=your_wandb_api_key +export DAYTONA_API_KEY=your_daytona_api_key +``` + +Prepare datasets (run from the SkyRL root): + +```bash +uv run examples/train_integrations/harbor/prepare_harbor_dataset.py \ + --dataset open-thoughts/CodeContests +``` + +Launch training: + +```bash +bash examples/train_integrations/harbor/run_codecontest.sh +``` + +See the [SkyRL README](https://github.com/NovaSky-AI/SkyRL) for full setup instructions.