This repository contains datasets associated with the paper "Rethinking Generalization in Reasoning SFT: A Conditional Analysis on Optimization, Data, and Model Capability".
The research investigates the factors influencing cross-domain generalization in Large Language Models (LLMs) during reasoning-focused supervised fine-tuning (SFT) with long chain-of-thought (CoT) data.
Optimization Dynamics: Cross-domain… See the full description on the dataset page:
https://huggingface.co/datasets/jasonrqh/NuminaMath-20k.
