
Hugging Face released a runnable recipe that trains harness-specific LoRA adapters using RL. On Qwen3.8-27B, a custom adapter boosts Terminal-Bench performance from 39% to 55% while cutting token usage roughly in half. The setup uses TRL, OpenEnv, and Harbor to train inside the actual agent loop.
Worth trying if you are stuck on a specific model and harness combination. The main catch is the infrastructure overhead: you need to manage Docker images and proxy logging to ensure the trainer learns from exact production tokens.
more: https://huggingface.co/blog/harness-adapters-at-scale
my other side: ivo.blogspot.com
Leave a Reply