EdgePro · LFM2.5-Audio-1.5B-JP — F-SOAIP nursing notes (on-device)

Listens to a spoken Japanese care handoff (申し送り) and writes a structured F-SOAIP nursing record as JSON — fully on-device, no cloud, no patient data leaving the building.

Fine-tuned from LiquidAI/LFM2.5-Audio-1.5B-JP with LoRA (≈0.85% of the weights, r=16). The adapter is merged into the base, so this repo loads exactly like a full model — no PEFT step needed.

On a held-out set it matches a frontier-cloud cascade (small edge model transcribes → GPT-5.5 structures) on note quality, while running entirely offline.

⚠️ Trained on synthetic data and graded by an LLM judge against a reference note. This is a research artifact, not a clinically validated medical device.

Usage

pip install "liquid-audio>=1.3" torchaudio   # Python >=3.12; CUDA recommended

import torchaudio
from liquid_audio import ChatState, LFM2AudioModel, LFM2AudioProcessor

REPO = "schang-jp/edgepro-lfm2-audio-v6-lora-long"   # base: "LiquidAI/LFM2.5-Audio-1.5B-JP"
processor = LFM2AudioProcessor.from_pretrained(REPO, device="cuda").eval()
model = LFM2AudioModel.from_pretrained(REPO, device="cuda").eval()

SYSTEM = (
    "あなたは介護・看護記録の専門家です。この申し送り音声を聞いて、"
    "F-SOAIP形式の6セクション記録をJSONで作成してください。"
    "キーは focus, subjective, objective, assessment, intervention, plan の6つで、"
    "値はすべて日本語の文字列です。余計なキーは出力しないでください。"
)

wav, sr = torchaudio.load("handoff.wav")            # ~30s mono care handoff
chat = ChatState(processor)
chat.new_turn("system"); chat.add_text(SYSTEM); chat.end_turn()
chat.new_turn("user"); chat.add_audio(wav.to(model.device), sr); chat.end_turn()
chat.new_turn("assistant")

# IMPORTANT: collect token ids, then decode the WHOLE sequence once.
# Decoding per-token corrupts multibyte kanji that span byte-BPE tokens (→ U+FFFD).
ids = [int(t.item()) for t in model.generate_sequential(**chat, max_new_tokens=1024) if t.numel() == 1]
note = processor.text.decode(ids)
print(note)   # {"focus": "...", "subjective": "...", ..., "plan": "..."}