maxylium/chess-qwen35-4b-rl

Chess Qwen 3.5 4B (v2 Zero-Hallucination Champion)

An advanced, end-to-end trained Chess Reasoning & Move Generation LLM built on Qwen 3.5 4B, fine-tuned and aligned on an AMD Instinct MI300X (192GB VRAM):

1. Curated Full SFT: 202,785 high-quality chess reasoning traces (candidate captures, checks, PV analysis, and best UCI moves).
2. Zero-Hallucination GRPO v2 RL: Online GRPO with severe punitive illegal move penalties (-15.0) and +2.0 legal move baseline incentives on 62,794 dynamic self-play, endgame, and GM puzzle positions.
3. Logit-Constrained High-Elo Engine: 100% legal move adherence across all multi-ply tournament play.

Evaluation Benchmarks

1. Static Accuracy & Engine Alignment (478 Held-Out Positions)

Metric	SFT Base	RL v1	RL v2 Champion
Format Compliance	100.0%	100.0%	100.0%
Legal Move Rate	90.8%	91.8%	92.7%
Stockfish Top-1 Match	32.2%	31.8%	31.0%
Stockfish Top-3 Match	48.5%	48.1%	47.7%
Mean Centipawn Loss (ACPL)	329.1 cp	328.2 cp	330.7 cp

2. Multi-Elo Ladder Benchmark Results

Opponent Tier	Engine Elo	Games	W / D / L	Model Score %	Retries Used
Beginner Baseline (200–600 Elo)	200–600	5	1/1/3	30.0%	33
Stockfish 1350	1350	5	0/0/5	0.0%	22
Stockfish 1600	1600	5	0/0/5	0.0%	53
Stockfish 1850	1850	5	0/0/5	0.0%	40
Max Stockfish (Skill 20 / 3000+ Elo)	3000+ (Max)	3	0/0/3	0.0%	18

How to Play

import chess, re
from unsloth import FastModel

model, tokenizer = FastModel.from_pretrained("maxylium/chess-qwen35-4b-rl")
FastModel.for_inference(model)

board = chess.Board()
prompt = f"You are a chess engine. Analyze the board inside <think></think> tags and output the best UCI move inside <move></move> tags.\nFEN: {board.fen()}\nBoard:\n{board}"

inputs = tokenizer.apply_chat_template([{'role': 'user', 'content': prompt}], add_generation_prompt=True, return_tensors='pt').to('cuda')
output = model.generate(inputs, max_new_tokens=256, temperature=0.2)
text = tokenizer.decode(output[0], skip_special_tokens=True)
print(text)