Upload README.md with huggingface_hub

README.md

@@ -10,28 +10,46 @@ tags:
 
 Neural networks (NNUE) for the [GaiaChess](https://github.com/jromang/gaiachess) UCI chess engine.
 
-**Architecture (GaiaNet)**: 12 king buckets × 768 PST + filtered threat features (~41K inputs) → 640 → CReLU + pairwise → 16 → 32 → 1, with 8 output buckets. Trained with [Bullet](https://github.com/jw1912/bullet) on self-play data.
-
 Files are append-only: a network file, once published, is never modified — engine releases pin an exact file name.
 
+## Architectures
+
+Two incompatible generations live here. Each network carries a 16-byte integrity footer
+(architecture hash + content hash) and the engine refuses, loudly, a network whose
+architecture is not the one it was built for — so a file from one generation cannot be fed
+to an engine built for the other.
+
+**GaiaNet T1** (`gaianet-t1-*`, `gaianet-t2-*`, `gaianet-t3*`) — 12 king buckets × 768 PST
+inputs + filtered threat features (~41K inputs) → 640 → CReLU + pairwise → 16 → 32 → 1,
+8 output buckets.
+
+**GaiaNet T2b** (`gaianet-t2b-*`) — the current generation. 12 king buckets × 768 PST inputs
+(i16 rows) plus 64,368 auxiliary features in i8 rows: 4,560 pawn pairs and 59,808 filtered
+piece-interaction (threat) features, accumulated into a 1024-wide layer → paired SCReLU →
+16 → 32 → 1, with 8 material-based output buckets and a PSQT head. 87 MB per file.
+
+Trained with [Bullet](https://github.com/jw1912/bullet) (T1) and the in-house trainer (T2b),
+on Leela T80 positions and self-play data.
+
 ## Networks
 
-| File | Engine version | Notes |
-|------|----------------|-------|
-| `gaianet-t1-1000.bin` | never shipped | archived for reproducibility: this network generated part of the training data used by its successors |
-| `gaianet-t2-1000.bin` | 4.2.0 – 4.2.5 | previous default network |
-| `gaianet-t3b-1000.bin` | unreleased | +12.0 ±9.5 Elo vs t2 (SPRT, 2588 games) |
-| `gaianet-t3e-1100.bin` | unreleased | t3b fine-tuned for 100 more superbatches at lr/10 |
-| `gaianet-t3e-1100b.bin` | unreleased (default in `main`) | current default network: same weights as t3e-1100 plus a 16-byte integrity footer (architecture + content hashes, verified by the engine at load; older engine binaries ignore it) |
+| File | Architecture | Engine version | Notes |
+|------|--------------|----------------|-------|
+| `gaianet-t1-1000.bin` | T1 | never shipped | archived for reproducibility: this network generated part of the training data used by its successors |
+| `gaianet-t2-1000.bin` | T1 | 4.2.0 – 4.2.5 | previous default network |
+| `gaianet-t3b-1000.bin` | T1 | unreleased | +12.0 ±9.5 Elo vs t2 (SPRT, 2588 games) |
+| `gaianet-t3e-1100.bin` | T1 | unreleased | t3b fine-tuned for 100 more superbatches at lr/10 |
+| `gaianet-t3e-1100b.bin` | T1 | unreleased | last T1 default: same weights as t3e-1100 plus the 16-byte integrity footer |
+| `gaianet-t2b-e013.bin` | **T2b** | unreleased (default in `main`) | current default network. Fine-tune of the T2b line on 62 GB of Leela T80 2024 positions, 60 superbatches at lr 1e-4 cosine. +25.4 ±9.7 Elo vs its parent (SPRT) |
 
 ## Usage
 
 Networks are embedded into the engine at build time:
 
 ```bash
-curl -L -o nets/gaianet-t3e-1100b.bin \
-  "https://huggingface.co/jromanghf/gaiachess-networks/resolve/main/gaianet-t3e-1100b.bin"
-MODEL=nets/gaianet-t3e-1100b.bin RUSTFLAGS="-C target-cpu=native" \
+curl -L -o nets/gaianet-t2b-e013.bin \
+  "https://huggingface.co/jromanghf/gaiachess-networks/resolve/main/gaianet-t2b-e013.bin"
+MODEL=nets/gaianet-t2b-e013.bin RUSTFLAGS="-C target-cpu=native" \
   cargo build --release --features nnue
 ```