yb2698 merges a week of LoRA adapters into one super-adapter, boosting Qwen coding scores by up to 20 points
yb2698 released an experimental "super-adapter" for Qwen3.8-27B: LoRA adapters trained over one week for different agent harnesses (toolchains) are concatenated along the rank dimension into one big LoRA, letting the model gain each harness's improvements without switching adapters. Measured results rose markedly; weights and training recipe are open-sourced, making this low-cost multi-task adapter merging approach worth watching.
Confirmed
- The merge covers all harness adapters trained this week (three in total), concatenated along the rank dimension, each at ½ weight, into a higher-rank LoRA.
- Rationale: the author notes the three harness adapters' weight updates are nearly orthogonal (cosine similarity 0.01), so concatenation merges them with almost no loss.
- Results (thinking on, xhigh, baseline → super-adapter): average across 6 benchmarks rose from 55.5% to 63.5%; on SWE-Bench Pro (hard 51), Hermes Agent went from 45% to 61%, OpenCode from 47% to 67%, pi held at 74%; OpenCode gained about 20 points; tool calls dropped by up to 73% (Terminal-related).
- Weights are live on Hugging Face (harness-hub/Qwen3.8-2…), accompanied by a full training-recipe blog and open-source training tutorial.
Why It Matters
- Demonstrates an extremely simple LoRA merging path: as long as adapters for different tasks are near-orthogonal, concatenation integrates multiple capabilities with near-zero loss, eliminating multi-adapter management and switching overhead.
- In the blog and follow-up posts, the author emphasizes that readers can train their own adapters for their tasks, models, and harnesses and merge them the same way, forming a reusable personal "harness-hub" workflow.
2026-11-20 ~ 2026-11-20 · 9 related posts
- Episode 1: Hugging Face's Harness Adapters: Cheap LoRA+RL Tune Small Models per Agent Framework(2026-11-06, 18 posts)
- Episode 2: Hugging Face's Harness Adapters: RL-trained LoRA doubles Qwen3.8-27B on Pi coding agent(2026-11-17, 16 posts)
- Episode 3: Hugging Face Releases RL-Trained OpenCode LoRA Adapter for Qwen3.8-27B(2026-11-18, 7 posts)
- Episode 4: yb2698 merges a week of LoRA adapters into one super-adapter, boosting Qwen coding scores by up to 20 points(2026-11-20, 9 posts)
Primary sources
- [source] Super-adapter merges weekly harness LoRAs for Qwen3.8-27B, lifting SWE-Bench Pro up 20 points — yb2698 · 2026-11-20
- Super-adapter lifts SWE-Bench Pro by 10+ points, cuts tool calls up to 73% — yb2698 · 2026-11-20
- Recipe released: train your own harness adapter for any model on your own tasks — yb2698 · 2026-11-20
- Super-adapter lifts Qwen3.8-27B agents from 55.5% to 63.5% mean across 6 pairs — yb2698 · 2026-11-20
- [source] Merging LoRAs via rank concatenation: orthogonal adapters combine losslessly — yb2698 · 2026-11-20
- [source] Qwen3.8-27B super-adapter weights are up on Hugging Face with full recipe — yb2698 · 2026-11-20
- One rank-96 LoRA boosts Qwen3.8-27B across 3 agent harnesses, 55.5% to 63.5% mean — HauntingMoment · 2026-11-20