AISA-Framework/AISA-AR-FunctionCall-FT
AISA-Framework/AISA-AR-FunctionCall-FT is a 0.3 billion parameter model built on FunctionGemma (Gemma 3 270M) by AISA-Framework. It is specifically fine-tuned for reliable Arabic structured tool calling, converting natural Arabic requests into executable API calls. This model excels at enabling integration between language models and external tools within Arabic agentic systems, significantly reducing parse failures and improving function name accuracy compared to its baseline.
Loading preview...
Overview
AISA-AR-FunctionCall-FT is a 0.3 billion parameter model developed by AISA-Framework, built upon the FunctionGemma (Gemma 3 270M) architecture. It is specifically fine-tuned for Arabic function calling and tool invocation, translating natural Arabic requests into structured API calls compatible with the FunctionGemma format. This model is a key component of the AISA (Agentic AI Systems Architecture) initiative.
Key Capabilities
- Arabic natural language to structured API calls: Converts diverse Arabic inputs into actionable tool calls.
- Multi-dialect Arabic understanding: Supports 5 Arabic dialects (MSA, Gulf, Egyptian, Levantine, Maghrebi).
- Enhanced reliability: Achieves a significant reduction in parse failure rate (from 87% to <1%) and improved function name accuracy (from 8% to 65%) compared to the baseline model.
- Broad domain support: Optimized for 8 real-world domains including Travel, Utilities, Islamic services, Weather, Healthcare, Banking & finance, E-commerce, and Government services.
Training and Evaluation
The model was trained using a data-centric fine-tuning pipeline on the AISA-AR-FunctionCall dataset, which comprises over 50,000 samples across 27 tool schemas. Evaluation on a held-out test set demonstrated substantial improvements in structured tool-call matching and argument extraction.
Intended Use Cases
- Arabic AI assistants
- Tool-based agents
- Structured API orchestration
- Arabic enterprise automation
- Research on multilingual tool calling