Qwen-2.5-Instruct

The 'Qwen-2.5-Instruct' dataset/benchmark contains high-quality expert human solutions used to evaluate and improve the reasoning capabilities of large language models through a method that transforms these solutions into in-distribution reasoning traces.

Papers using Qwen-2.5-Instruct (3)

Mechanistic origins of catastrophic forgetting: why RL preserves circuits better than SFT?2026Didactic to Constructive: Turning Expert Solutions into Learnable Reasoning2026Training Superior Sparse Autoencoders for Instruct Models2025