Harness adapters: figures
Figures embedded in the harness adapters blog post.
One base model, an adapter per harness
Four task types
One base model, many adapters
From sessions to a gradient step
From a shared trajectory to an RL task
One trajectory, three patterns
Five Repo2RLEnvs tasks, file by file
A reverse proxy between the harness and the model
Interleaving the task mixture
Synchronous vs asynchronous RL (from the AsyncGRPOTrainer post)
Training inside the harness
Training correctness
Tool calls per solved rollout
Task groups per training step
Results per harness
Merging nearly orthogonal adapters
Asynchronous training loop, per harness