Distilling Stockfish with One Billion Positions (blog.lukesalamone.com)

A researcher announced the release of the Gigafish dataset, the largest open state-value chess dataset comprised of 1 billion positions extracted from a total of 3.9 billion chess positions. This significant achievement allows for the training of neural networks to predict the evaluation of chess positions, potentially enhancing AI models' performance in chess without relying solely on search depth, as demonstrated in prior research by DeepMind. The model built from this dataset features a hybrid architecture with 79 million parameters, combining 6 ResNet blocks and 16 Transformer blocks. It achieved a directional accuracy of 92.97%, meaning it successfully identifies the winning side correctly in 19 out of 20 cases. The dataset's construction involved meticulous processes to deduplicate positions and annotate each with Stockfish evaluations. By leveraging this innovative dataset and architecture, the research may pave the way for more accurate AI-driven chess evaluations and strategies, marking a notable advancement in the field of AI and machine learning applications in gaming.