FULL STORY

HF's Harness Adapters: From Live Training to Release

Hugging Face's harness-hub team live-streamed the training of LoRA+RL harness adapters to make smaller models better agents, then officially released the first adapters, including Qwen3.8-27B-pi-adapted, roughly doubling solve rates.

2026-11-06 ~ 2026-11-17 · 2 episodes · 22 posts

Episode 1 · Hugging Face's Harness Adapters: Cheap LoRA+RL Tune Small Models per Agent Framework (2026-11-06, 18 posts)

Hugging Face's harness-hub team (yb2698, ariG23498) released and live-streamed the training of "harness adapters": instead of modifying agent harnesses, a small LoRA adapter is trained for each model×harness pair with verifiable-task RL, executed inside unmodified real coding environments (NousResearch's hermes, pi, opencode). Covered models include LiquidAI's LFM2.5-24B-A2B (MoE, 2B active), LFM3-1.7B and thewebAI's TwIL-8B. Lysandre (Liquid AI) and Aritra (HF) introduced the work, and HF published an accompanying blog.

Confirmed

  • Single-GPU LoRA fine-tuning of LFM3-1.7B: after a 20-step smoke run, on 150 held-out tasks (paired vs. baseline, 95% CI) the opencode harness gained 9.4 ± 6.3 points pass@1 (69%→79%)
  • Compute-saving setup: one vLLM per model, one trainer per harness, each mounting LoRA adapters on shared weights; the three LFM3-1.7B trainers share one vLLM, and evaluating 1350 agent sessions cost only $5
  • Reward design: reward = task verifier score + 0.1 × (1 − calls/15), applied only to correct rollouts; ablations show a 0.3 coefficient hurts accuracy while 0.1 cuts tool calls by 11–17% with no accuracy loss
  • Environment synthesis: seeded with public agentic trajectories from the HF Hub, converting only trajectories containing verifiable units of work into tasks, yielding Repair/Build/Extend/QA environments, difficulty-stratified via LFM2.5-24B self-assessment to offset the old environments' data-analysis bias
  • Training uses the safeagents anti-cheating scheme with all tool calls proxied and audited to prevent reward hacking; environments and training pipeline are open-sourced alongside the stream

Unconfirmed

  • The exploration directions (new RL tasks from trajectories, adding the opencode harness, cross-domain transfer) remain goals of this round; full live-stream results pending
  • Hesamation's prediction that 2027 will be the year of extreme model–harness co-design is personal speculation without evidence

Why it matters

  • The work shows small models can gain clear improvements on a specific agent framework via extremely cheap ($5-scale) targeted RL + LoRA, offering a reproducible open-source paradigm for model–framework co-optimization; the open-sourced environment synthesis and anti-cheating tooling lower the barrier for community replication

Episode 2 · Harness Adapters launch: LoRA doubles Qwen3.8-27B coding scores (2026-11-17, 4 posts)

Hugging Face and harness-hub released Harness Adapters, RL-trained LoRA adapters that tune models like Qwen3.8-27B to specific agent harnesses. On the Pi harness, solve rates doubled, with SWE-bench Pro rising from 29.4% to 62.7% and tool calls cut by 55%.