RL Training Integration¶
OpenSage-ADK exposes evaluation benchmarks as RL rollout targets. This guide covers two integrations, AReaL and slime. A third adapter, miles, is registered in src/opensage/evaluation/rl_adapters/adapters/ but does not yet have a dedicated guide.
| Framework | Rollout Engine | Trainer | Benchmark Coverage |
|---|---|---|---|
| AReaL Training | SGLang (TP=2) | FSDP (DP=2), GRPO | SeCodePLT |
| slime Training | SGLang in a SLIME container | Megatron-LM, SLIME | SeCodePLT, mock |
Both setups wrap OpenSage agents as the rollout environment and thread trajectories back to the trainer.