Key Info

An open experiment reports finished harness-adapter training for Liquid AI's LFM2.5-24B and LFM3 and thewebAI's Twil-8B, optimized for the pi, hermes and opencode harnesses, using more than 1,900 RL steps on Hugging Face jobs.

Highlights

  • LFM3 adapters show up to 43% better tool efficiency and a 15% accuracy gain on a held-out data-analysis test set.
  • Twil-8B adapters show up to 37% improvement in tool-call efficiency.
  • 9 adapters total are released with GGUF versions ready to use.
  • Reward-hacking attempts were blocked for all models via safeagents, and the attempts themselves were open-sourced.