distil-labs/distil-lfm25-home-assistant

Overview

The distil-lfm25-home-assistant model is a compact yet powerful 0.35 billion parameter language model, fine-tuned by distil labs from the LiquidAI/LFM2.5-350M base model. Its core function is to enable multi-turn smart home control via tool calling, translating natural language commands into structured function calls for devices like lights and thermostats. This model is particularly adept at handling conversational adjustments and sequences of commands.

Key Capabilities & Performance

This fine-tuned model demonstrates exceptional performance in its specialized domain:

• High Tool Call Equivalence: Achieves an impressive 96.7% tool call equivalence, significantly surpassing the 63.2% of its base model and even outperforming a 120B parameter teacher model by 4.6 percentage points.
• Robust ROUGE Score: Boasts a ROUGE score of 99.4%, indicating high accuracy in generating relevant and coherent responses.
• Efficient Multi-Turn Conversations: Designed to manage complex, multi-turn interactions where users refine or chain commands.
• Optimized for Embedded Deployment: Its small size and high accuracy make it ideal for deployment on smart home hubs and IoT gateways, enabling sub-second latency for voice commands without cloud round trips. It supports runtimes like ONNX, Ollama, vLLM, and llama.cpp.

Training Details

The model was trained using Supervised Fine-Tuning (SFT) with LoRA on the distil-labs/distil-smart-home dataset, leveraging the distil labs platform. It utilizes the LFM2.5 tool calling format, marked by <|tool_call_start|> and <|tool_call_end|> tags.