distil-lfm25-voice-assistant

A fine-tuned version of LiquidAI/LFM2.5-350M for multi-turn banking voice assistant via tool calling, trained using the distil labs platform.

This model routes customer requests to the correct banking function (check balance, transfer money, cancel card, report fraud, etc.) while extracting required parameters. With 14 distinct functions, complex slot types, and ASR transcription artifacts in the input, this is the hardest task in our benchmark suite.

Results

Metric	Teacher (120B)	LFM2.5-350M Base	LFM2.5-350M Tuned
Tool Call Equivalence	96.95%	34.5%	95.9%
ROUGE	97.55%	74.1%	99.0%

The tuned model comes within 1.1 points of the 120B teacher on the hardest task in our benchmark suite.

Training Details

Parameter	Value
Base model	LiquidAI/LFM2.5-350M
Teacher model	GPT-oss-120B
Task type	Multi-turn tool calling (closed-book)
Training data	distil-labs/distil-voice-assistant-banking
Training method	SFT with LoRA
Platform	distil labs

Training Progress

Epoch	Tool Call Equivalence
0 (base)	34.5%
1	86.8%
2	92.4%
3	95.9%
4	95.4%

Usage

This model uses the LFM2.5 tool calling format with <|tool_call_start|> and <|tool_call_end|> tags:

<|tool_call_start|>[function_name(arg1="value1", arg2=42)]<|tool_call_end|>

Deployment

This model targets on-device banking workflows where customer data cannot leave the device perimeter, deployable on mobile NPUs via ONNX or on Apple Silicon laptops via MLX for branch-level kiosk applications. It also works with Ollama, vLLM, llama.cpp, or any inference runtime that supports Safetensors.

Blog Post

For the full writeup, see: Fine-Tuning Liquid's LFM2.5: Accurate Tool Calling at 350M Parameters

License

This model is licensed under the LFM Open Model License v1.0.