lfm2-herdr-gguf

Herdr expert — LFM2-350M (GGUF)

A merged LoRA fine-tune of LiquidAI/LFM2-350M, trained by the herdr-liquid-finetune repo to be an expert on the Herdr terminal multiplexer (25 Herdr tool-calling ops, plus off-topic refusal). Exported to GGUF for llama.cpp so it runs as a single binary — no Python needed.

The adapter lives at agney/lfm2-herdr-lora; this repo is the same weights merged into the base and quantized.

Files

Run (llama.cpp)

# 1. build llama.cpp (once): cmake -B build -G Ninja && cmake --build build
./build/bin/llama-cli \
  -m lfm2-herdr-Q4_K_M.gguf \
  -p "$(cat prompt.txt)" \
  -st -n 128 --temp 0

prompt.txt is the same system env + Herdr tool schemas the training prompt used — build it with the fine-tune repo's tokenizer.apply_chat_template(tools=...) (or copy the 25 herdr_tools.SCHEMAS into the prompt). The model answers in the native <|tool_call_start|>[name(k=v, ...)]<|tool_call_end|> syntax; parse it with the repo's eval_lfm2.parse_calls.

For an OpenAI-style HTTP endpoint use llama-server with the same model, and send your chat request with tools=.

Measured accuracy (pinned 120-row holdout, v8)

Scored with scripts/eval_gguf.py (same prompt + parse_calls as the eval_lfm2.py transformers eval). All quantizations keep 100% off-topic; the diff is exact-call and a couple of arg-grounding rows.

Lossless: F16 and Q8_0 reproduce the bf16 adapter exactly; the merge + GGUF conversion introduces no behavior change. Q4_K_M costs ~4 pts (mostly the pane_current / pane_split arg-grounding rows the bf16 model already struggles with). Pick Q8_0 for accuracy, Q5_K_M for the best size/accuracy balance, Q4_K_M only if size is the priority.

Notes

• Narrow specialist: plans the 25 Herdr ops and refuses off-topic prompts; does not do general chat, code, or reasoning.
• GGUF is a storage format — you still need the llama.cpp binaries to run it, and you still format the prompt/tool list yourself. It removes the Python model-loading layer, not the inference harness.
• License: MIT (project weights). The base model remains under its own license.