---
license: mit
tags:
  - chess
  - transformer
  - recurrent
  - policy
  - endgame
  - pytorch
library_name: pytorch
---

# 99M endgame specialist (squares64)

Same **99M** squares64 architecture as
[`avewright/chess-transformer-100m-squares64`](https://huggingface.co/avewright/chess-transformer-100m-squares64),
finetuned on `<14`-piece positions (one-hot best first move) from
[`Lichess/chess-position-evaluations`](https://huggingface.co/datasets/Lichess/chess-position-evaluations)
via [`avewright/lichess-endgame-bestline`](https://huggingface.co/datasets/avewright/lichess-endgame-bestline).

This file is **`latest.pt` at endgame-FT step 4500** (2026-09-13 23:41 UTC).
Train loss ~1.4738. Frozen holdout hard CE ~1.2928.

Not the generalist incumbent, the puzzle expert, or the Syzygy expert.
Not [`avewright/endgame-dataset`](https://huggingface.co/datasets/avewright/endgame-dataset) (SF19 MultiPV harvest).

## Training

- Warm start: public 99M `latest.pt` (weights only), then full resume.
- Split: position-hash 80/20 (seed 276). Frozen piece-stratified val 8192.
- One-hot PV1 (`soft_alpha=0`). Pieces 2–13.
- Polar-NorMuon, bs=528. Best disk ckpt at upload: step 4500.

## Files

- `latest.pt`
- `step_004500.pt`
- `model_config.json`
- `train.log`
- `pack.json`