ptrdvn/kakugo-apc

Kakugo Levantine Arabic dataset

[Paper] [Code] [Model]

A synthetically generated conversation dataset for training in Levantine Arabic.

This dataset contains synthetic conversational data and translated instructions designed to train Small Language Models (SLMs) for Levantine Arabic. It was generated using the Kakugo pipeline, a method for distilling high-quality capabilities from a large teacher model into low-resource language models. The teacher model used to generate this dataset was openai/gpt-oss-120b.

For Kakugo in other languages, check out the model and dataset collections.

Creation Methodology

This dataset was created using the automated Kakugo pipeline described in our paper. Full details of how this dataset was created (and how you can make a dataset in your own chosen language) can be found on our Github repo.

1. Synthetic Data Generation

We prompted a teacher model (GPT-OSS 120B) to generate diverse prompts in Levantine Arabic using three strategies:

• Topic-Based: Prompts derived from a tree of general and language-specific topics (e.g., local culture, history, daily life).
• Scenario-Based: Prompts based on realistic user scenarios where an AI assistant would be useful (e.g., "planning a trip," "explaining a concept").
• Context-Based: Prompts generated by feeding the teacher model random text snippets from HuggingFaceFW/fineweb-2 in Levantine Arabic and asking it to perform tasks like summarization, translation, or QA based on that text.

For every generated prompt, the teacher model produced a response. Crucially, we captured the teacher's reasoning traces (chain-of-thought) to help the student model learn how to think, not just what to say.

2. Instruction Translation

To bolster general instruction-following capabilities, we sampled high-quality English instructions from the BAAI/Infinity-Instruct (7M_core) dataset.

• These were translated into Levantine Arabic using the teacher model.
• Strict filtering was applied: conversations were discarded if the translated length was disproportionate (indicating hallucination or failure) or if the formatting was broken.

Usage & Limitations

• Thinking Mode: The dataset includes specific system prompts that trigger "thinking mode." Thinking mode is trained for the data we have reasoning traces for - only our synthetically generated data. When training on this data, the model learns to output <think> tags containing reasoning steps only when prompted.
• Synthetic Nature: While the teacher model is highly capable, this data is synthetic or machine-translated. This dataset is NOT PERFECT!

Credit

This model was trained by @ptrdvn

If you use this dataset, please cite the Kakugo paper:

@article{devine2026kakugo,
  title={Kakugo: Distillation of Low-Resource Languages into Small Language Models},
  author={Devine, Peter and Sanni, Mardhiyah and Adilazuarda, Farid and Loizaga, Julieta Gil and Haddow, Barry},
  journal={arXiv preprint arXiv:2601.14051},
  year={2026}
}