xxang/AStar-Thought-V2-OpenR1-Math-3k

AStar-Thought-V2-OpenR1-Math-3k This dataset is the training data used for the A-Thought-V2* models presented in A*-Thought-V2: Efficient Latent Reasoning via Geometric Dynamics of LLM. It is built from OpenR1-Math-3k and compressed using the A*-Thought-V2 pipeline, which interleaves explicit text reasoning with implicit latent tokens guided by the geometric dynamics of the LLM. Each example includes a system prompt, latent conversations, original conversations, segmented steps… See the full description on the dataset page: https://huggingface.co/datasets/xxang/AStar-Thought-V2-OpenR1-Math-3k.