HF team live-trains harness adapters: LoRA + RL to make small models better at agentic tasks
yb2698 · x · 2026-11-06
The Hugging Face harness-hub team kicked off a new round of live-streamed "harness adapters" training: one small LoRA adapter per model×harness pair, trained with RL on verifiable tasks inside unmodified agent harnesses (pi, opencode, hermes).
- Models: LiquidAI's LFM2.5-24B-A2B (MoE, 2B active, 200 steps), LFM3-1.7B (400 steps), plus experimental TwIL-8B (100 steps)
- Tasks: SmolDataEnvs and 230 traj2envs tasks; each model's three trainers share one vLLM
- safeagents anti-reward-hacking: a proxy inspects every tool call; undeclared tools, unknown hosts or high-severity rule hits are blocked and score 0, lower-severity ones are logged as suspicious
- Eval: held-out pass@1 scheduled per checkpoint under its own harness; smoke runs show adapters cut tool calls while keeping or improving accuracy
Live dashboard streams train reward, blocked/suspicious events and all metrics.
More from coding & agent
- Dev shows multi-agent crew auto-creating a Black Friday discount code end-to-end for $0.10 — pramodk73 · 2026-11-06
- Developer says Codex wrecked his server after a reboot request, costing $10,000+ — burny_tech · 2026-11-06
- Exo: Berkeley's agent runtime that edits its own harness, cutting model cost 35-61% — dair_ai · 2026-11-06
- Symbolic belief revision plus geometric projections lifts RAG reranking to 0.481 nDCG@10 — letandrewcook · 2026-11-06
- Free open-source AI Engineering from Scratch course teaches math-first — HowDevelop · 2026-11-06
- Charter: A Lean 4 DSL for Formally Verified Rules Governing AI Agents — QuixiAI · 2026-11-06