agney/lfm2-herdr-lora

LFM2-Herdr Expert (LoRA adapter)

A PEFT LoRA adapter over LiquidAI/LFM2-350M, fine-tuned to be an expert on the Herdr terminal multiplexer: given a natural-language request, it emits the correct Herdr tool call (or refuses off-topic prompts). This is a narrow specialist — it plans the 25 Herdr operations, not a general chat/code/reasoning model.

Loading

Load it on top of the base model with peft:

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

tok = AutoTokenizer.from_pretrained("LiquidAI/LFM2-350M")
model = AutoModelForCausalLM.from_pretrained(
    "LiquidAI/LFM2-350M", dtype=torch.bfloat16, device_map="auto")
model = PeftModel.from_pretrained(model, "agneym/lfm2-herdr-lora").eval()

prompt = tok.apply_chat_template(
    [{"role": "system", "content": "HERDR_ENV=1\nworkspace=w1\ntab=w1:t1\npane=w1:p1\ncwd=/home/repo\nagent kind=hermes"},
     {"role": "user", "content": "split my pane"}],
    tools=..., tokenize=False, add_generation_prompt=True)
ids = tok(prompt, return_tensors="pt").to(model.device)
out = model.generate(**ids, max_new_tokens=192, do_sample=False)
print(tok.decode(out[0][ids.input_ids.shape[1]:], skip_special_tokens=False))

The model answers in native <|tool_call_start|>[name(k=v, ...)]<|tool_call_end|> syntax. Load the tool schemas from `reference/herdr_schemas.json` in the training repo.

Evaluation

Scored on the pinned 120-row holdout (runs/results/eval_v8_holdout.json, seed 42, strictly disjoint from training), all 25 tools represented:

exact-call requires the tool name and arguments to match the label (key-order-insensitive, pane_split normalized to current=true).

Training

• —Base: LiquidAI/LFM2-350M (bf16, gradient checkpointing), T4/L4.
• —Data: 804 rows in dataset.jsonl (98 off-topic, 12.2%), system-prompt rotation over 8 contexts so grounding comes from the prompt, not a memorized w1:p1 / /home/repo constant.
• —LoRA: r=16, alpha=32, dropout 0.05, targets q_proj/k_proj/v_proj/w1/w3/w2 (the LFM2 MLP projections are w1/w3/w2, not gate/up/down_proj; do NOT target out_proj, which is shared with Lfm2ShortConv).
• —SFT: 12 epochs, batch 1, grad-accum 8, lr 1e-4, cosine schedule, loss masked to assistant tokens only, best-val checkpoint.

Limitations

• —Fails some novel paraphrases ("give me a new pane on the right" → pane_create(Direction=...); "where am i?" under-calls).
• —pane_split and pane_current argument grounding is below 100% on the holdout.
• —Does not do general chat/code/reasoning; it plans the 25 Herdr ops and refuses off-topic prompts.

License

MIT. The full pipeline (dataset generation, training, eval) is in `herdr-liquid-finetune`.