HF boosts Qwen3.8-27B from 39% to 65% on SWE-bench Pro with RL and one small LoRA
TheTuringPost · x · 2026-11-19
Hugging Face's Clement Delangue team shared a concrete recipe to fit an open model to your coding agent: RL plus a small LoRA lifts Qwen3.8-27B from 39% to 65% on SWE-bench Pro (hard 51), and from 39.5% to 55.8% on Terminal-Bench 2.1.
- Keep the harness, adapt the model: the same model behaves differently across harnesses; instead of switching harnesses, train a small LoRA adapter per harness and swap it in behind your existing agent (first one targets the pi harness).
- Train inside the real, unmodified harness: the harness runs as-is in its own sandbox per rollout; training uses exactly the tokens the harness saw — no re-tokenisation, no simplified copy.
- Reward correctness, then efficiency: task verifier correctness plus a tool-efficiency bonus (weight 0.1).
More from coding & agent
- Stripe engineer bans herself from slides, uses AI to build interactive presentations in 30 minutes — hazelcough · 2026-11-19
- Microsoft reveals Autopilot architecture with OpenClaw running in an untrusted sandbox — steipete · 2026-11-19
- Forward-deployed engineers: tie AI projects to a business metric before writing code — alex_verem · 2026-11-19
- Modem admits its AI PM agent flopped, rebuilds platform to work with your own agents — jevon · 2026-11-19
- Running an agent fleet across 15 repos: the bottleneck turned out to be the human — yi_ding · 2026-11-19
- Local MCP server lets Claude search your PDFs, old emails, voice memos and photos offline — IndependentBook5854 · 2026-11-19