wraith1337/ravager-net-ckpt

KYOKASUIGETSU

An NNUE evaluation net for the Ravager chess engine, trained from scratch (not derived from any other net) with bullet on free Google Colab T4 GPUs.

Status: training in progress. Cosine learning-rate schedule over 600 superbatches (100M positions each). kyokasuigetsu-latest.bin is a snapshot from superbatch 24 and is NOT a finished net. Per-superbatch checkpoints (weights + optimiser state + loss log) are under ckpt/.

• Architecture: 768x5 mirrored king buckets -> 640 (x2 perspectives) SCReLU -> 1x8 output buckets (material count). QA=255, QB=64, eval scale 400. Same binary layout as Leorik's 640HL net (4,937,024 bytes), loadable by Ravager via EvalFile.
• Training data: Renegade training data (public domain), files 260312 and 260217, filtered (no checks, no tactical positions, no early plies, |eval| <= 10000, 15% random skip). Result weight 0.4.
• Recipe: AdamW, batch 16384, LR 1e-3 cosine to 2.4e-6. See ravager_net.rs and run.py.
• Strength has not been measured yet. Do not assume it beats any existing net.