SupraLabs Releases 5-Million-Sample Reasoning Dataset for Small Language Models
SupraLabs has released reasoning-corpus-4K-5M-v1, a large-scale dataset containing 5 million samples designed for training small language models. The dataset features chain-of-thought traces and is formatted in ChatML. This release provides a massive, open-source resource that enables developers to fine-tune smaller, resource-efficient language models with advanced reasoning capabilities. It helps democratize high-quality AI training by reducing the computational barrier for complex task performance. All samples in the dataset are kept within a 5k sequence length, making them ideal for supervised fine-tuning (SFT). Each entry includes metadata such as the source repository ID, token length, user prompt, thought trace, and final assistant response.
## BACKGROUND
Chain-of-thought (CoT) prompting is a technique that improves AI reasoning by breaking down complex problems into intermediate, step-by-step logical stages. ChatML (Chat Markup Language) is a structured format used to represent multi-turn conversations, helping models distinguish between system, user, and assistant roles.