A LoRA fine-tune of LiquidAI/LFM2.5-230M for on-device to-do tool-calling. Given a short natural-language utterance and the user's current to-do list, pomo-1 emits a single structured tool call to create, update, or delete a to-do. Built to run locally on Apple Silicon via MLX.

This is a task-specific model, not a general assistant. It does one thing: turn an utterance + a small list of existing to-dos into one JSON tool call.

Intended use

In scope: single-turn to-do CRUD intent → one tool call, on-device.
Out of scope: multi-turn dialogue, reasoning, general chat, code, or any task the base model is not recommended for (advanced math, code generation, creative writing). Inherits the base model's limits.

Tools

tool	arguments
`create_todo`	`title` (str), `due` (str \| null) — ignores the current list
`update_todo`	`target` (str), `title` (str \| null), `due` (str \| null)
`delete_todo`	`target` (str)
`none`	`{}` — emitted when the referenced to-do is not in the list

For update_todo / delete_todo, target is a verbatim copy of one item in the provided list. If the referenced item is absent, the model emits none.

Prompt format

The current to-do list (0–5 items) is injected into the prompt. Training and inference must use this exact layout:

Todos:
- <todo 1>
- <todo 2>

User: <utterance>

An empty list renders as Todos:\n(none). Output is a JSON string, e.g.:

{"name":"delete_todo","arguments":{"target":"Book the moving truck"}}

Usage (MLX)

python

from mlx_lm import load, generate
from mlx_lm.sample_utils import make_sampler

model, tokenizer = load("<YOUR_HF_REPO>")  # e.g. sabeshbesh/pomo-1

prompt = "Todos:\n- Book the moving truck\n- Water the front yard plants\n\nUser: delete moving truck task"
tokens = tokenizer.apply_chat_template(
    [{"role": "user", "content": prompt}],
    add_generation_prompt=True, tokenize=True,
)
out = generate(model, tokenizer, prompt=tokens, max_tokens=96,
               sampler=make_sampler(temp=0.0))
print(out)  # {"name":"delete_todo","arguments":{"target":"Book the moving truck"}}

Greedy decoding (temp=0.0) is recommended for deterministic tool calls. The base model's general-purpose defaults are temperature 0.1, top_k 50, repetition_penalty 1.05.

Training

Base: LiquidAI/LFM2.5-230M (LFM2 hybrid: 14 layers, 8 gated short-conv + 6 GQA).
Method: LoRA SFT via mlx_lm.lora.
Adapter: rank 16, scale 16, dropout 0.05, applied to all 14 layers; mask_prompt: true.
Data format: legacy {"prompt", "completion"} JSONL; completion is a JSON-string tool call.
Hardware: Apple Silicon (MLX).

Evaluation

Held-out validation (n=300), best checkpoint, greedy decoding:

metric	score
parse rate (valid JSON)	1.000
tool-name accuracy	0.997
argument exact-match	0.563
full match (name + args)	0.563

Metric definitions: parse rate = fraction of outputs that are valid tool-call JSON; tool-name accuracy = correct tool selected; argument exact-match = arguments exactly correct; full match = both correct.

Known limitation of this eval: these figures were measured on a dataset variant where the current to-do list was not included in the prompt, which makes correct target selection for update/delete structurally impossible in many cases — the argument/full-match scores are a floor, not a ceiling. No baseline comparison against the prior model is included. Treat these numbers as provisional.

License

Governed by the base model's license, lfm1.0. See the base model license.

Citation

Base model:

bibtex

@article{liquidAI2026230M,
  author  = {Liquid AI},
  title   = {LFM2.5-230M: Built to Run Anywhere},
  journal = {Liquid AI Blog},
  year    = {2026},
  note    = {www.liquid.ai/blog/lfm2-5-230m}
}