Skip to content

v0.2.0

Pre-release
Pre-release
Compare
Choose a tag to compare
@sgrvinod sgrvinod released this 17 Dec 18:18
· 11 commits to main since this release
56028b7

New release of Chess Transformers – v0.2.0

  • New model: CT-EFT-20, which predicts the source (From) and destination (To) squares for the best half-move. Significantly outperforms earlier models, while also being the smallest.
  • Standardized vocabulary: All models and datasets will hereon use a standard in-built vocabulary, eliminating the need for keeping track of separate vocabulary files, and allowing models to be used across datasets, and datasets across models, with ease.
  • More stringent evaluation: Fairy Stockfish, used to evaluate trained models, is now configured to run on more threads and with a much larger hash table, making it a more challenging opponent. Evaluation results for earlier models have been recomputed.
  • Easier installation: Only up to a maximum of 3 environment variables need to be set when training or evaluating a model.
  • Changelog added: A changelog now tracks changes associated with each release, in detail.

Still in initial development. There may be some bugs.