0x3/functiongemma-finetuned-g1-multilingual
The 0x3/functiongemma-finetuned-g1-multilingual model is a 0.3 billion parameter FunctionGemma variant, fine-tuned by 0x3 to convert natural language into structured robot action and emotion function calls. It supports six languages (English, Chinese, Japanese, French, German, Spanish) with 98% accuracy. This model is optimized for low-latency robot control, achieving approximately 59ms inference time on NVIDIA Jetson AGX Thor, making it suitable for real-time robotic applications.
Loading preview...
Overview
This model, developed by 0x3, is a fine-tuned version of Google's FunctionGemma 270M, specifically designed for converting natural language commands into structured robot actions and emotions. It operates with a compact 0.3 billion parameters and a context length of 32768 tokens.
Key Capabilities
- Multilingual Support: Processes commands in 6 languages: English, Chinese, Japanese, French, German, and Spanish.
- High Accuracy: Achieves 98% accuracy in translating natural language to robot functions.
- Low Latency: Delivers responses in approximately 59ms on NVIDIA Jetson AGX Thor, crucial for real-time robotic interactions.
- Function Calling: Generates specific robot actions like
shake_hand,face_wave,hands_up,stand_still,show_hand,do_payment, anddown_payment. - Emotion Generation: Can trigger robot emotions such as
happy,sad,excited,confused,curious, andthink. - Efficient Decoding: Utilizes constrained decoding with 2 forward passes, resulting in an ~18x speedup compared to standard autoregressive generation.
Training Details
The model was fine-tuned using LoRA (rank 8, alpha 16) on a dataset of approximately 6,000 examples, including 545 English and 5,450 multilingual entries. The multilingual data was generated using the Claude API to ensure diverse and natural phrasing across languages, rather than literal translations. Training was conducted for 3 epochs with a learning rate of 2e-4 and a batch size of 4 (effective 16).