A new dataset called Scaffold CoT has been released, containing approximately 4 million examples and 3 billion tokens. This dataset is designed to improve the reasoning and response accuracy of smaller language models (under 5 billion parameters) by providing a structured thinking framework. The framework consists of three consistent sections: 'Inventory', 'Interaction', and 'Execution', which helps models focus on content rather than formatting, reducing errors and increasing coherence. Scaffold CoT also includes diverse categories and depth tiers to enhance topic-specific performance and adaptable reasoning. AI
IMPACT Aims to improve the reasoning capabilities of smaller language models, potentially making them more accessible and reliable for specific tasks.
RANK_REASON Release of a new dataset for LLM training. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
