Modelpublic

Paulescu/LFM2.5-Audio-1.5B-OHF-Voice-GGUF

sourceHugging Faceotherupdated 5mo agoView on Hugging Face

5likes194downloads

Model Card

LFM2.5-Audio-1.5B-OHF-Voice

Fine-tuned from LiquidAI/LFM2.5-Audio-1.5B on Paulescu/OHF-Voice-audio-20260504 to map spoken Home Assistant voice commands directly to function calls. Part of the Liquid Cookbook voice-assistant example.

Output format (no system prompt; the function set is baked into the weights):

HassStartTimer|$minutes=5|$name=oven
HassLightSet|$area=bedroom|$brightness=70
HassGetCurrentTime

Files

llama-liquid-audio-server requires four GGUFs to run inference:

file	description	source
`LFM2.5-Audio-1.5B-OHF-Voice-Q4_0.gguf`	Language model backbone	fine-tuned in this repo
`mmproj-LFM2.5-Audio-1.5B-OHF-Voice-Q4_0.gguf`	Audio encoder + projector	fine-tuned in this repo
`vocoder-LFM2.5-Audio-1.5B-OHF-Voice-Q4_0.gguf`	Audio decoder (unused for function-calling)	copied from upstream
`tokenizer-LFM2.5-Audio-1.5B-OHF-Voice-Q4_0.gguf`	Tokenizer / speaker file	copied from upstream

The runners/ folder bundles llama-liquid-audio-server and llama-liquid-audio-cli binaries for macos-arm64, ubuntu-x64, ubuntu-arm64, and android-arm64, built from llama.cpp PR #18641.

Reproduce the eval

bash

git clone https://github.com/Liquid4All/cookbook
cd cookbook/examples/voice-assistant
uv sync
# point configs/finetuned-Q4_0.yaml at this repo and run:
uv run python scripts/eval.py --config configs/finetuned-q8.yaml