Amortized Planning with Large-Scale Transformers: A Case Study on Chess
Anian Ruoss, Grégoire Delétang, Sourabh Medapati, Jordi Grau-Moya, Li Kevin Wenliang, Elliot Catt, John Reid, Lewis, Cannada A., Veness, Joel, Genewein, Tim · arXiv (Cornell University) · 2024
This dataset was prepared for the experiments in the Chess Reasoner project. It is derived from ChessBench, the large-scale chess benchmark introduced by Ruoss et al. in Amortized Planning with Large-Scale Transformers: A Case Study on Chess. The original ChessBench dataset contains chess positions from Lichess games annotated with legal moves and Stockfish 16 value estimates. This release extends the selected ChessBench-derived positions with additional Stockfish MultiPV line annotations. These annotations are used to train and evaluate transformer-based chess models that generate candidate continuations before selecting a final move. The archive contains the processed dataset used in our experiments. The corresponding code repository contains the preprocessing scripts, configuration files, random seeds, and a small sample dataset.