RL-Trained Hermes LoRA on Qwen3.8-27B Adds +13.7 Pts on SWE-Bench Pro While Using Fewer Tokens
yb2698 · x · 2026-11-19
A developer released a GGUF LoRA adapter for Qwen3.8-27B, trained with RL (TRL's AsyncGRPOTrainer + OpenEnv) on real Hermes Agent sessions in an HF sandbox, with safeagents checks on every tool call.
- Results: SWE-bench Pro (hard 51) 29.4% → 43.1% (+13.7 pts); Terminal-Bench 2.1 44.7% → 58.8% (+14.1 pts, p=0.029)
- Efficiency: under 40% of the base model's turns on commonly solved tasks, 2× fewer tokens
- No-thinking mode: matches the base model at xhigh reasoning effort while generating 3.5–4.4× fewer tokens per task — a big win for local deployment
The author openly flags caveats: the thinking comparison holds only under a 131k window with one attempt per task; the GGUF file wasn't re-evaluated through llama.cpp; scores aren't comparable to the official Qwen3.8-27B card. All evaluation trajectories are linked.
Related event: Nous Releases RL-Trained Hermes Adapter for Qwen3.8-27B(2 posts)→
More from coding & agent
- Qwen3.8-27B + Hermes adapter lifts SWE-bench Pro to 43% with 4x fewer tokens — yb2698 · 2026-11-19
- Every on Claude Fable 5.2: Human Thinking at the Start Matters More as AI Builds More — every · 2026-11-19
- Semantic caching LLM answers by embedding similarity is unsafe — 138-pair test shows why — Narrow_Side_7778 · 2026-11-19
- Can ComfyUI refine existing mocap animations? A motion-to-motion workflow question — _Tito5_ · 2026-11-19
- Mycelium: an open-source framework that turns analytical project knowledge into a living dataset — arjunrajlab · 2026-11-19
- harness-hub open-sources RL-trained LoRA adapters tuned per agent harness, recipe coming Friday — yb2698 · 2026-11-19