milsoni201/lfm25-350m-axstream-matcher

axstream-matcher (LFM2.5-350M fine-tune)

The instant-tier matcher for axstream — a streaming action language for computer-use agents. Given a library of known voice-command templates and a spoken utterance, this model answers one constrained question in ~100ms on Apple Silicon: which known command is this, and what are the variable words?

input  (system): the user's command library — id, description, slots, examples
input  (user):   "uh can you open apple music for me"
output:          {"template": "open_app", "slots": {"app": "apple music"}}

It never plans, never sees the screen, never invents actions — it recognizes. Misses return "none", which routes the utterance to a full LLM. Slot values are copied verbatim from the utterance by contract (train- and serve-time enforced), so the matcher cannot act on words the user didn't say.

Files

file	what
lfm25-350m-axstream-v2-Q4_K_M.gguf	v2 — current. Trained on variable 10–40-template libraries, sibling families, deliberate look-alike traps, and description-only templates.
lfm25-350m-axstream-Q4_K_M.gguf	v1 (kept for reproducibility).
adapter_model.safetensors	v1 LoRA adapter over LiquidAI/LFM2.5-350M.

v2 results

v2 targets the failure modes observed running v1 in production: accuracy collapse past ~15 in-context templates, wrong matches on plausible unknown commands ("open apple maps" with no maps template present), and never matching templates that lack examples. Evaluated through the real serving path (llama-server + JSON-schema constrained decoding), v1 vs v2 on the same suites:

gate	v1	v2
exact canonical examples @ 25-template libraries	96.0%	100%
wrong-match rate on deliberate-rival non-commands	23.3%	4.8%
accuracy at 40-template libraries	73%	91%
library-size sweep spread (10→40 templates)	11 pts	4 pts
paraphrase holdout (generated by a model family unused in training)	—	94.0%
description-only (example-less) templates	~0%	92.2%
latency p50 (M-series, Q4_K_M, prefix-cached)	~110ms	~120ms

In axstream, v2 is paired with an embedding prerank (top-10 shortlist, measured recall@3 = 100% on the live library), which keeps the effective context in the 10-candidate sweet spot at any library size.

v1 results

Evaluated through the real serving path (llama.cpp llama-server + JSON-schema constrained decoding) on a 521-example held-out test set of dictated-speech commands over dynamic 8–14-template libraries, including compound-sentence negatives, near-miss app names, and chit-chat:

metric	base LFM2.5-350M	this model
end-to-end correct (template + slots)	47.3%*	94.2%
template accuracy	53.2%*	96.0%
slot values exact (given template)	53.3%*	97.5%
refusal recall on non-commands	55.0%*	94.3%
wrong-match rate on non-commands	45.0%*	5.7%
latency p50 (M-series, Q4_K_M, prefix-cached)	104ms	114ms

*base measured on the round-1 test set (387 examples); this model's numbers are from the harder superset. Trained in 3.7 minutes on one RTX 4090 (~$0.10).

Usage

Serve the GGUF with llama.cpp and constrain the output with a JSON schema over your template ids (grammar-constrained decoding makes invalid output impossible). The exact prompt builder and schema this model was trained against are open source: axstream/tiny.py.

llama-server -m lfm25-350m-axstream-v2-Q4_K_M.gguf --port 8791 -ngl 99 -c 8192 --no-webui

(v2 was trained with prompts up to 3072 tokens — give the server headroom with -c 8192.)

Or as the matcher inside axstream directly — see the quickstart.

Training

LoRA (r=32, α=64) on all projections including LFM2's conv-block in_proj / out_proj and MLP w1/w2/w3, per the official unsloth LFM2.5 recipe. Prompts are byte-identical to the serving format; response-only masking; trained on one rented RTX 4090. The dataset and generation pipeline are not released.

v1: ~2.6k synthetic dictation-style examples with machine-validated labels (verbatim-slot checks, cross-model judge filtering), 3 epochs at lr 5e-5, seq 2048. ~4 minutes, ~$0.10.

v2: ~5.5k train rows over a harder curriculum — every context library sampled at 10–40 templates (median 23, shuffled); ~12 authored families of near-identical sibling commands trained co-present; ~580 deliberate-rival negatives (a plausible command whose true template is absent while a confusable same-verb template is present → gold none); 400 positives whose true template renders description-only; paraphrase refresh across six spoken style axes including whisper-transcript artifacts. Generation by one model family, judging by two others, and the paraphrase holdout by a fourth used nowhere in training. 2 epochs at lr 5e-5, seq 3072. Well under $1.

Limitations

• English, macOS-command domain, dictated-speech register.
• v2 is in-distribution for 10–40-template contexts; beyond that, pre-filter (axstream preranks to a top-10 shortlist with a local embedding model).
• v2 traded a little recall for its much lower wrong-match rate: some preamble-heavy phrasings ("open the browser and search for …") over-refuse and route to the LLM tier — the safe direction.
• Slot values are verbatim by contract, so tasks whose spoken word maps to a different UI value (say "times", click the button named "Multiply") are out of scope for the matcher by design.
• This is a recognizer. Pair it with guards before acting on its output — in axstream, replays verify targets against the live screen and anything destructive is risk-gated.

License and attribution

This model is a derivative of LiquidAI/LFM2.5-350M by Liquid AI, Inc., modified by merging a LoRA adapter fine-tuned on synthetic voice-command data and quantizing to GGUF (Q4_K_M). Distributed under the LFM Open License v1.0 (see LICENSE).

Note: under LFM Open License v1.0 §5, commercial use by entities (including affiliates) with ≥ $10M annual revenue is not licensed. This restriction travels with the weights and cannot be relicensed away.

Not affiliated with or endorsed by Liquid AI, Inc.