TuwaiqAcademy/AISA-AR-FunctionCall-Think
TuwaiqAcademy/AISA-AR-FunctionCall-Think is a compact 270M-parameter Arabic function-calling model, fine-tuned from google/gemma-3-270m. It is designed to generate an Arabic reasoning trace ( tag) before emitting a structured tool call, supporting five Arabic dialects. This model serves as the official baseline for Track B (Reasoning-Augmented Function Calling) of the AISA-ArabicFC shared task, excelling in tool selection and providing a lightweight starting point for Arabic agentic experiments.
Loading preview...
AISA-AR-FunctionCall-Think: Reasoning-Augmented Arabic Function Calling
This model, developed by Tuwaiq Academy, is the official Track B baseline for the AISA-ArabicFC shared task at ArabicNLP 2026. It is a compact 270M-parameter model, fine-tuned using LoRA from google/gemma-3-270m, specifically designed for Arabic function calling.
Key Capabilities & Features
- Reasoning-Augmented Function Calling: Generates a short Arabic
<think>reasoning trace before emitting a structured tool call, a core requirement for Track B of the shared task. - Multidialectal Arabic Support: Handles user queries in Modern Standard Arabic (MSA) and five major Arabic dialects: Gulf, Egyptian, Levantine, and Maghrebi.
- Task-Specific Optimization: Fine-tuned on the AISA-ArabicFC reasoning data to accurately decide if a function call is needed, select the correct function, and extract structured arguments.
- Strong Baseline Performance: Achieves an FnAcc of 0.982 and an ArgEM of 0.541, significantly outperforming zero-shot GPT-4o on the AISA-ArabicFC evaluation metrics, demonstrating the value of task-specific training.
- Low Hallucination Rate: Exhibits a 0.000 hallucination rate on negative (no-tool) queries.
Intended Use Cases
- Shared Task Baseline: Serves as a reference baseline for participants in the AISA-ArabicFC shared task, providing a reproducible starting point.
- Lightweight Arabic Tool-Use: Offers a resource-efficient model for initial experiments and development in Arabic agentic systems and tool-use applications, particularly within the defined AISA-ArabicFC schema.
Limitations
- Schema-Specific: Optimized for the 27-tool, 8-domain AISA-ArabicFC schema and its specific prompt format; performance on arbitrary tools or free-form chat is not guaranteed.
- Single-Turn/Single-Call: Designed for single-turn interactions and single function calls, not multi-tool or multi-turn dialogues.
- Argument Extraction Imperfections: While strong, argument extraction (ArgEM 0.541) can still show errors in areas like date normalization, numeric typing, and dialectal phrasing.