endgame expert model card
Browse files
README.md
ADDED
|
@@ -0,0 +1,40 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: mit
|
| 3 |
+
tags:
|
| 4 |
+
- chess
|
| 5 |
+
- transformer
|
| 6 |
+
- recurrent
|
| 7 |
+
- policy
|
| 8 |
+
- endgame
|
| 9 |
+
- pytorch
|
| 10 |
+
library_name: pytorch
|
| 11 |
+
---
|
| 12 |
+
|
| 13 |
+
# 99M endgame specialist (squares64)
|
| 14 |
+
|
| 15 |
+
Same **99M** squares64 architecture as
|
| 16 |
+
[`avewright/chess-transformer-100m-squares64`](https://huggingface.co/avewright/chess-transformer-100m-squares64),
|
| 17 |
+
finetuned on `<14`-piece positions (one-hot best first move) from
|
| 18 |
+
[`Lichess/chess-position-evaluations`](https://huggingface.co/datasets/Lichess/chess-position-evaluations)
|
| 19 |
+
via [`avewright/lichess-endgame-bestline`](https://huggingface.co/datasets/avewright/lichess-endgame-bestline).
|
| 20 |
+
|
| 21 |
+
This file is **`latest.pt` at endgame-FT step 4500** (2026-09-13 23:41 UTC).
|
| 22 |
+
Train loss ~1.4738. Frozen holdout hard CE ~1.2928.
|
| 23 |
+
|
| 24 |
+
Not the generalist incumbent, the puzzle expert, or the Syzygy expert.
|
| 25 |
+
Not [`avewright/endgame-dataset`](https://huggingface.co/datasets/avewright/endgame-dataset) (SF19 MultiPV harvest).
|
| 26 |
+
|
| 27 |
+
## Training
|
| 28 |
+
|
| 29 |
+
- Warm start: public 99M `latest.pt` (weights only), then full resume.
|
| 30 |
+
- Split: position-hash 80/20 (seed 276). Frozen piece-stratified val 8192.
|
| 31 |
+
- One-hot PV1 (`soft_alpha=0`). Pieces 2–13.
|
| 32 |
+
- Polar-NorMuon, bs=528. Best disk ckpt at upload: step 4500.
|
| 33 |
+
|
| 34 |
+
## Files
|
| 35 |
+
|
| 36 |
+
- `latest.pt`
|
| 37 |
+
- `step_004500.pt`
|
| 38 |
+
- `model_config.json`
|
| 39 |
+
- `train.log`
|
| 40 |
+
- `pack.json`
|