---
license: mit
library_name: pytorch
tags:
- chess
- transformer
- policy
- soft-targets
- stockfish
---

# Chess Transformer 200M (compact vocab + soft MultiPV FT)

Finetune of [`avewright/chess-transformer-200m-latest`](https://huggingface.co/avewright/chess-transformer-200m-latest)
onto **compact move vocab (1968)** using local Stockfish MultiPV soft targets
([`avewright/exp186-sf-multipv-2m`](https://huggingface.co/datasets/avewright/exp186-sf-multipv-2m)).

## Checkpoint
- File: `best_model.pt` (exp188 `best.pt` snapshot)
- Architecture: 16L / 1024d / 16H, ~204M params, fused encoder, spatial policy, 3-class WDL
- Vocab: **compact 1968** (`MOVE_VOCAB_VERSION=compact`)
- Soft holdout top-1 at upload: **33.82%** (step 1500 of resume; training may still be running)

## Load
```python
import os
os.environ["MOVE_VOCAB_VERSION"] = "compact"
import torch
from huggingface_hub import hf_hub_download
from chess_transformer_factory import build_model, DEFAULT_200M_CONFIG

path = hf_hub_download("avewright/chess-transformer-200m-compact-soft", "best_model.pt")
model = build_model(DEFAULT_200M_CONFIG)
ckpt = torch.load(path, map_location="cpu", weights_only=False)
model.load_state_dict(ckpt["model_state_dict"], strict=False)
model.eval()
```

## Training notes
See `TRAINING_STATUS.md` in this repo.

Uploaded: 2026-07-10T02:02:52.143564+00:00