2026-03-13
Changed
- Training pipeline substantially refactored: all text preprocessing (including regex chunking and pretokenization) is now performed by a parallelized Rust-side workflow. This reduces training time by over 80% and cuts memory usage by approximately 50%.