agi-noobs/chess-sft-100k-chess-lora-20251230-010330

Chess LoRA - Fine-tuned for Chess Move Prediction

This is a LoRA adapter fine-tuned on chess positions to predict the best move.

Training Details

Parameter	Value
Base Model	Qwen/Qwen3-4B-Instruct-2507
Dataset	agi-noobs/chess-sft-100k
Training Steps	2,812
Unique Examples	90,000
Total Examples	89,984
Test Samples	10,000
Epochs	1
Batch Size	32
LoRA Rank	128
LoRA Alpha	32
LoRA Params	264,241,152 (6.57% of model)
Total Model Params	4,022,468,096
Learning Rate	0.0004905250962784578
LR Schedule	cosine
Total Tokens	26,862,473
Budget	N/A
WandB	chess-lora-20251230-010330

Evaluation Results

Metric	Value
Best Move Accuracy	14.0%
Legal Move Rate	99.0%
ACPL (mean)	323.7
ACPL (median)	171.0
Good Moves (<50 CPL)	31.3%
Inaccuracies (50-100)	9.6%
Mistakes (100-300)	18.7%
Blunders (300+)	40.4%
Puzzle Accuracy	25.0%

Prompt Template

You are an expert chess player. Here is the position in FEN format:
{{ FEN }}

Legal moves: {{ legal_moves_san }}

Select the best move. Keep your thinking to 2 sentences or less, then output your chosen move in SAN first, then convert it to UCI.

CRITICAL: Your move must always be one of the moves from legal moves list. With no spaces in between.

An example of a valid move for this position is <uci_move>{{ first_legal_move }}</uci_move>. 
Format:
<rationale>brief thinking</rationale>
<san>Nf3</san>
<uci_move>g1f3</uci_move>

Response Template

<rationale>{{ thinking }}</rationale>
<san>{{ best_move_san }}</san>
<uci_move>{{ best_move_uci }}</uci_move>

Sample Positions

Sample 1

FEN: rnbq1rk1/pp2ppbp/3p1np1/2p5/2PPP3/2N1BP2/PP4PP/R2QKBNR w KQ - 0 7 Best Move: g1e2

Sample 2

FEN: 4n3/6b1/p2pN1kp/2pPpN2/2P1P1p1/P4q2/2Q2P2/6K1 b - - 6 37 Best Move: f3a3

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("YOUR_USERNAME/YOUR_REPO")
tokenizer = AutoTokenizer.from_pretrained("YOUR_USERNAME/YOUR_REPO")

prompt = "<|im_start|>user\n[Your chess prompt here]<|im_end|>\n<|im_start|>assistant\n"
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=100)
print(tokenizer.decode(outputs[0]))

Training Infrastructure

Trained using Tinker by Thinking Machines Lab.