Makeen-AraFC/glaive-tool-use-reasoning-ar

Arabic Glaive Tool-Use Reasoning

Arabic translation and augmentation of the Glaive Function Calling data for research on Arabic function calling, tool selection, argument generation, and tool-call verification.

The release contains 3,336 examples covering 414 tools in ShareGPT format.

Dataset Structure

Each example contains:

{
  "tools": [...],
  "conversations": [...]
}

• —`tools`: candidate tool declarations, including names, descriptions, parameter names, types, and required parameters.
• —`conversations`: ShareGPT-style messages containing the system context, Arabic user request, and target function call.

Tool and parameter identifiers are preserved from the source data.

Reasoning Traces

The original Glaive examples do not contain reasoning traces. The Arabic Glaive reasoning traces were therefore generated during the translation and augmentation process.

For each example, the generation prompt was provided with:

• —the user request;
• —the target function call;
• —the already-translated tool declaration.

The model was instructed to generate a concise 2–3 sentence Arabic reasoning trace explaining the tool selection and argument assignment.

Thus, the think traces in this dataset are synthetically generated supervision, not translated reasoning traces from Glaive.

Translation

Tool declarations and conversations were translated using dedicated prompts.

The translation process preserves:

• —tool names;
• —parameter names;
• —parameter types;
• —API identifiers and enum values;
• —JSON structure.

Natural-language descriptions and user requests are translated into Arabic. Argument values originating from the user request are aligned with the translated request.

Statistics

Intended Use

The dataset is intended for research on:

• —Arabic function calling;
• —tool selection and argument generation;
• —tool-call verification and correction;
• —Arabic instruction tuning;
• —multilingual tool use.

Source

This dataset is derived from the Glaive Function Calling v2 dataset.

Citation

TODO