chessjev-minilm-v1

Chess-specialized one-pass option scorer. MiniLM-L6 encodes a FEN/feature context and each legal-move option; a small attention head returns N logits in one pass (jevlike Choice).

This is not TypeSafe Jev. It is not Elo-tested. Do not treat valid accuracy as playing strength.

Item	Value
Architecture	sentence-transformers/all-MiniLM-L6-v2 + option-attention head
Teacher	Local Stockfish 19, 40 ms/position
Data	3000 train / 400 valid / 200 test {context, options, label} JSONL
Train	Full finetune, Apple M1 MPS, 3 epochs, batch 8, lr 2e-5
Valid teacher-agreement	0.445 (16-option shortlist; random ≈ 0.06)
Valid tactic-agreement	0.667 (small held-out mate/tactic slice)
Code	github.com/gopalanj/chess-jev (MIT)
Encoder license	Apache-2.0 (MiniLM). ChessJev training code is MIT.

Files

• model.pt — encoder + option head (MiniLMOptionScorer)
• config.json — architecture hyperparameters

Use

# from the chess-jev repo
hf download gopalanj/chessjev-minilm-v1 --local-dir checkpoints/chessjev-minilm-v1
PYTHONPATH=src uvicorn chess_jev.server:app --host 127.0.0.1 --port 8765

from pathlib import Path
from chess_jev.minilm import load_minilm

model = load_minilm(Path("model.pt"))

App fallback if this file is missing: chessjev-byte-v1, then the heuristic scorer.