Scaffold CoT: A 4M Example Structured Reasoning Dataset for Small Models

Saraozte01 · reddit · 2026-08-25

The author released Scaffold CoT, a dataset of 4M examples and 3B tokens designed to improve free-form Chain of Thought (CoT) in small models (<5B params). It enforces a fixed structure (Inventory -> Interaction -> Execution) to standardize reasoning and reduce hallucinations.

Covering 18 domains and 798 subdomains including code, general knowledge, anti-hallucination, and tool use, the dataset allows models to focus on content rather than formatting. The rigid structure also enables regex validation of outputs, facilitating the training of reliable small models on consumer hardware.

Original post →

More from Research

Research channel →