Hugging Face livestreams open RL training: trajectory-built envs, safeagents, LoRA evals
yb2698 · x · 2026-11-06
Hugging Face will livestream its next RL harness-adapter training run, building accurate and efficient LoRA adapters for the hermes, pi and opencode coding harnesses on LiquidAI's LFM2.5-24B, LFM3 and TwIL-8B — with envs, reward-hacking attempts, eval results and the training protocol all open. Four enhancements over the last iteration:
- Repo2RLEnvs: new environments synthesized from open-source agent trajectories on the HF Hub
- safeagents: mitigates reward hacking, harness introspection and credential misuse
- Swapped Claude Code for opencode
- Eval optimizations: one GPU server can evaluate up to 10 LoRA checkpoints at once
More from coding & agent
- Intelligence now 'too cheap to meter': investor drops Codex for all-you-can-eat Claude at $100 — StewartalsopIII · 2026-11-06
- One plain-English prompt turns your phone into a 200-particle physics simulator — tak3sh8 · 2026-11-06
- Zep launches Agent Skills: compile verified agent runs into reusable skills — TheTuringPost · 2026-11-06
- Sentry's next MCP evolution: one CLI login powers both MCP and full CLI access — zeeg · 2026-11-06
- Box launches Box Content Retrieval with agentic platform and MCP support at BoxWorks — shashib · 2026-11-06
- Give your agent an interpreter, not tools: 19 min to 2 min on the same model — Itz_hany · 2026-11-06