Skip to main content

Tinker Compatibility

Docker-only execution

All commands run inside Docker containers. Use the provided scripts.

W8-RL includes a Tinker adapter for token-based RL training using the Tinker Cookbook API. It wraps EnvActor directly and runs inside ray-worker.

How it works​

  • EnvActor provides reset_external / step_external
  • TinkerEnvAdapter converts ObsRef to Tinker tokens
  • GRPO training uses group rollouts and per-group advantage centering

Run rollout (no training)​

Run the command below from the repo root in Docker:

./scripts/run_design2code_tinker.sh \
  --task-dir tasks --limit 3 \
  --policy tinker \
  --model Qwen/Qwen3-4B-Instruct-2507 \
  --episodes 1 \
  --max-tokens 2048 \
  --max-total-tokens 3000000

Run training​

See Tinker RL Training.

Next Steps​