LoRA Adapter for Qwen3.8-27B: Terminal-Bench 2.1 Rises From 39.5% to 55.8% With In-Harness RL
Hugging Face trained adapters for coding agents with reinforcement learning. In pi, the Terminal-Bench 2.1 success rate rises from 39.5% to 55.8%, with roughly half the tokens per task.