rasbt on harness vs model fine-tuning: Qwen3.8-27B hit 65% on SWE-bench Pro
rasbt · x · 2026-11-21
Sebastian Raschka joins a discussion on SWE-bench improvement paths. TheTuringPost cited HF's recipe of keeping the harness and adapting the model: Qwen3.8-27B went from 39% to 65% on SWE-bench Pro with RL plus one small LoRA. rasbt says he saw the post last week and was thinking about how much difference the harness makes and why fine-tuning for a harness matters — framing the real question as which side is cheaper to change for your setup.
Related event: Debate over whether LLMs can absorb agent harness as RL tuning pays off(3 posts)→
More from coding & agent
- Multica Ships Realtime Docs for 100+ Humans and Agents, Replacing Feishu Internally — jiayuan_jy · 2026-11-21
- Epic Ports 20,000+ Fortnite Blueprints to Verse in Two Weeks with AI, Expected to Take Years — jradoff · 2026-11-21
- Merchant blocks shopping agent, loses $1,100 sale as bot buys locally instead — Scobleizer · 2026-11-21
- Threat actor Storm-3358 uses Hermes agent framework with skills that train AI to evade SOC detection — cyb3rops · 2026-11-21
- Matt Pocock narrows his claim: up-front planning only for costly, large rework — mattpocockuk · 2026-11-21
- housecor: re-prompting from scratch now beats building it once the old way — housecor · 2026-11-21