Team livestreams open RL training of LoRA harness adapters for hermes, pi and opencode
yb2698 · x · 2026-11-06
yb2698 and ariG23498 are livestreaming RL training on Hugging Face of LoRA "harness adapters" for three coding harnesses (hermes, pi, opencode), built on LiquidAI's LFM2.5-24B, LFM3 and the recent TwIL-8B models. The idea: fine-tune LoRA adapters jointly on the harness and the target task using RL and open environments, making models more efficient and accurate in the pipeline. Everything — RL environments, reward-hacking attempts, eval results, training protocol — will be open-sourced.
More from coding & agent
- API dies mid-task: only Atomic Agent with local fallback aced the test, Codex just hung — Ok-Brush8377 · 2026-11-06
- First Terminal Bench 2.1 results: ~1,300-line Mini Harness lands mid-pack but fastest — PawelHuryn · 2026-11-06
- Devs struggle with latency and interruption handling in Gemini Live voice agents — ProfessionalTank1886 · 2026-11-06
- Provider-level verifiable tool call results are essential for LLM agents — Sauers_ · 2026-11-06
- BeepGrid: open-source MCP-backed layer for agents to talk across runtimes — Eoncarry · 2026-11-06
- WebGL-based lens simulations let you play with optics right in the browser — zzznah · 2026-11-06