OrdenWills/LFM2.5-350M-home-assistant-sft
The OrdenWills/LFM2.5-350M-home-assistant-sft is a 350 million parameter smart home automation model, fine-tuned from LiquidAI's LFM2.5-350M. It specializes in translating natural language commands into structured tool calls for controlling smart home devices like lights, doors, thermostats, and media players, with full awareness of current device states. This model excels at complex disambiguation, multi-device commands, and robust rejection of unsupported actions, achieving 97.8% accuracy across 59 categories of smart home interactions. Its primary strength lies in reliable, structured tool call generation for Home Assistant environments.
Loading preview...
Model Overview
OrdenWills/LFM2.5-350M-home-assistant-sft is a specialized 350 million parameter language model, fine-tuned from LiquidAI/LFM2.5-350M, designed specifically for smart home automation. It processes natural language commands and current device states to generate precise tool calls for controlling various home devices, including lights, doors, thermostats, TVs, fans, and speakers.
Key Capabilities
- Advanced Device Disambiguation: Intelligently resolves ambiguous commands (e.g., "Turn off the TV") by considering device states and room context.
- State Awareness: Handles commands based on current device states, such as turning off only devices that are currently on, or gracefully responding when actions are already satisfied.
- Complex Command Handling: Supports universal scope commands ("Off everything"), pronoun resolution, relative state clauses ("Close the door that is open"), and multi-device compound commands.
- Rigid Syntactic Action Triggers: Employs internal reasoning to reliably signal whether an action is required before generating tool calls, preventing JSON bleed.
- High Accuracy: Achieves an overall accuracy of 97.8% across 59 specific edge-case and logic-stress categories in smart home scenarios.
Training and Features
The model was fine-tuned on a highly augmented 164,000-example synthetic dataset (V14), incorporating "Positive-Only Think Traces" to reduce hallucination and "Compound Count Enforcement" to prevent truncation of multi-action requests. It uses a 10-tool schema for device control.
Use Cases
This model is ideal for developers building robust, natural language interfaces for Home Assistant or similar smart home platforms, particularly where precise, state-aware tool execution is critical. It's suitable for applications requiring reliable command interpretation and structured output for device control.