metga97/functiongemma-270m-ar-tooluse
metga97/functiongemma-270m-ar-tooluse is a 270 million parameter Gemma-based model fine-tuned for Arabic tool use and function calling across various dialects and domains. Developed by metga97, it is specifically designed to generate a single, structured tool call using FunctionGemma's native formatting when a tool is required. This model excels at accurately parsing tool names and arguments, making it suitable for integrating function-calling capabilities into Arabic-speaking applications.
Loading preview...
Model Overview
This model, metga97/functiongemma-270m-ar-tooluse, is a specialized fine-tuned version of google/functiongemma-270m-it (270 million parameters) focusing on Arabic tool use and function calling. It is designed to accurately identify when a tool is needed and generate a single, properly formatted tool call with structured JSON arguments, adhering to FunctionGemma's native tool formatting.
Key Capabilities
- Arabic Tool Use: Optimized for function calling in Arabic, supporting multiple dialects (Egyptian, Gulf, Levantine, MSA, Maghrebi) and domains (banking, e-commerce, healthcare, travel, etc.).
- Structured Output: Produces exactly one tool call in the format
<start_function_call>call:TOOL_NAME{ ...json args... }<end_function_call>when a tool is required. - High Accuracy: Achieves a Tool name accuracy of 0.9921 and a Key-F1 (avg) of 0.9925 on its evaluation dataset, with a low missed-call rate of 0.0064.
- Domain-Specific Performance: Demonstrates strong performance across various domains, with Strict EM scores ranging from 0.4652 (utilities) to 0.8653 (weather).
Usage Considerations
- Inference Setup: Requires specific tokenizer settings (
tokenizer.padding_side = "left",tokenizer.pad_token = tokenizer.eos_token) and passing tools viaapply_chat_template(..., tools=tools_list). - Limitations: May occasionally translate slot values into English; mitigations include stronger prompt constraints or post-processing. Some outputs may have parsing inconsistencies that could be improved with longer training or more negative examples.