LoRA adapter lifts Qwen3.8-27B to 63% on SWE-Bench Pro, 60% fewer tool calls
ben_burtenshaw · x · 2026-11-20
- yb2698 released a LoRA adapter for Pi harness that RL-trains Qwen3.8-27B into a much better coding agent.
- Results: SWE-Bench Pro (Hard 51 subset) pass@1 jumps from 29% to 63%; Terminal-Bench 2.1 pass@1 rises from 40% to 56%.
- Efficiency: 55-62% fewer tool calls per task.
- Built entirely on the Hugging Face stack (TRL + Transformers), using the new stable AsyncGRPOTrainer.
- The author calls this the start of an "era of harness-specific adapters", with more coming.
More from coding & agent
- What 1.5 Years of Building Agents Taught Me: It's All a While Loop, Tools, and Specs — khant_dev · 2026-11-20
- TwinLab-Bench tests AI scientists in agentic twins of real labs; best model scores just 45.0 — zju-community · 2026-11-20
- Building an AI phone agent: 2018's duct-tape pipeline vs 2026 — JafarNajafov · 2026-11-20
- Open-Source AI Agent Eigin Will Pay Contributors for Specs, Not Bounties — danielrothmann · 2026-11-20
- X creator shares how agents now handle scheduling, research and data analysis — superinvestor_43 · 2026-11-20
- Mubit + Neon: agent middleware cuts frontier-model tokens by 73% — JeremyCMorgan · 2026-11-20