Egypt's TokenAI Releases Open Dataset of 50,000 Reasoning Records

The dataset, called The Forge Reasoning, is the first public release in a planned family of datasets the group calls The Forge. It covers six domains: mathematics, science, programming, defensive cybersecurity, artificial intelligence and machine learning. TokenAI says the combination is meant to give models working across software, science and authorised security tasks a shared foundation.

The release matters most to small labs and university teams. High-quality reasoning data, with full worked solutions that have been checked, is expensive to write and validate, and many developers instead generate it from larger commercial models. A free, consistently structured set lowers that cost for groups building or fine-tuning their own systems, including in Arabic-speaking markets where local labs have fewer resources.

Every record has six fields: domain, topic, question, solution trace, answer and common pitfall. The last field names a typical error for each problem, so researchers get an example of how a solution can go wrong alongside the worked trace showing the correct path. TokenAI says the uniform format is intended to make the reasoning easier to inspect, evaluate and reuse.

The dataset is published on Hugging Face under a CC BY 4.0 licence covering TokenAI's original material, which allows reuse and adaptation with attribution.

TokenAI has already used the data to train Horus Cyber Nano 1.2, an upcoming open-weights multimodal reasoning model aimed at authorised security analysis, secure coding, visual inspection and defensive technical work. The developer has not given a release date for that model.

The organisation was founded in 2025 by Assem Sabry as a nonprofit foundation to make AI models, research and tools more accessible. It plans to add further datasets, domains and evaluation resources to The Forge but has not announced details of those releases.