bpmatt/LFM2.5-350M-Phone-Command-Extractor-GGUF

LFM2.5-350M Phone Command Extractor (GGUF, Q4_0)

What it does: given one phone action and what the person said, it fills in the action's details using the person's own words, or declines when the action doesn't fit. It is the second stage of a fully on-device voice assistant for Android: the EmbeddingGemma-300M Phone Command Router picks the action, this model extracts the values, and app code turns spoken values ("half six", "tomorrow", "20 minutes") into exact times and dates.

action: add_reminder(text, time, date)      user: "remind me in 20 minutes to check the oven"
model:  <|tool_call_start|>[add_reminder(text='check the oven', time='20 minutes')]<|tool_call_end|>

It is a LoRA fine-tune of LiquidAI/LFM2.5-350M, merged and quantised to Q4_0 (219 MB). It runs with llama.cpp on a phone CPU in about 0.3-0.4 s per request.

Files

File	What it is
extractor-20261012T225126Z-Q4_0.gguf	the model (llama.cpp GGUF, Q4_0; chat template with native tool calls included)
system_prompt.txt	the exact system prompt it was trained with (the fixed date line is part of the format)
tool_schemas.json	the 205 actions as the model sees them (names, descriptions, parameter descriptions)

Quick start (llama.cpp)

llama-server -m extractor-20261012T225126Z-Q4_0.gguf --jinja -c 4096

Then send an OpenAI-style chat request with one tool from tool_schemas.json, the system prompt from system_prompt.txt, the user's words, temperature: 0 and max_tokens: 96. The reply is either a tool call (tool_calls[0].function.arguments) or the text No call.

import json, requests
tools = {t["name"]: t for t in json.load(open("tool_schemas.json"))}
r = requests.post("http://127.0.0.1:8080/v1/chat/completions", json={
    "messages": [{"role": "system", "content": open("system_prompt.txt").read().strip()},
                 {"role": "user", "content": "remind me in 20 minutes to check the oven"}],
    "tools": [{"type": "function", "function": tools["add_reminder"]}],
    "temperature": 0, "max_tokens": 96}).json()
print(r["choices"][0]["message"].get("tool_calls") or r["choices"][0]["message"]["content"])

Behaviour

• Copies values as spoken ("half six", "next Friday", "20 minutes", "fifty-five percent"): converting them into exact times, dates and numbers is left to deterministic code, which is easier to test and never invents a value.
• Leaves out values that weren't said instead of guessing.
• Says "No call." when the action doesn't fit the request or a required value is missing, so the app can ask.

Results

On 901 held-out requests (each given the action chosen by the router), the extracted call is exactly right (action and every value) 76.6 % of the time. Inside the full assistant (router + this model + safety rules + a confirmation step), 66 % of requests are completed correctly with no questions asked, 83 % are completed or settled with a single yes/no, and 4.6 % lead to a wrong action.

Training

LoRA fine-tune on about 7,000 examples in the model's native tool-call format: synthetic phone commands written for this project (every one of the 205 actions, with examples where the call must be declined), plus requests adapted from public intent datasets. Every example's values were checked to appear verbatim in the request and to pass the app's own value parser; held-out sets were checked for near-duplicates.

Limitations

• English only; one action per request (multi-step requests are split by the app).
• Expects the system prompt in system_prompt.txt, including its fixed date line, and one tool at a time.
• Spoken values need a resolver (relative times and dates are resolved by app code, not by the model).

Licence

A derivative of LFM2.5-350M by Liquid AI, distributed under the LFM Open License v1.0 (LICENSE), including its Commercial Use limitation (Section 5).