---
language:
- en
license: apache-2.0
tags:
- chess
- reinforcement-learning
- supervised-fine-tuning
- lora
- trl
- transformers
- qwen
pipeline_tag: text-generation
base_model: Qwen/Qwen3-0.6B
datasets:
- Lichess/chess-puzzles
---

# dopamineaddict/qwen3-0.6b-chess-sft-v1

This model is a LoRA fine-tuned variant of `Qwen/Qwen3-0.6B`, trained to output chess moves in UCI format wrapped in `<uci_move>...</uci_move>` along with a short `<rationale>...</rationale>` explanation.  
It was created as a baseline submission model for the AIcrowd Global Chess Challenge 2025.

## Intended use

- Intended for chess move selection in settings where the input includes:
  - A FEN position.
  - The side to move.
  - A list of legal moves in UCI.
- Intended output format:
  - `<uci_move>{legal_uci_move}</uci_move>`
  - `<rationale>{one-sentence explanation}</rationale>`

## How to use

Example (Transformers):

```
from transformers import AutoTokenizer, AutoModelForCausalLM
import torch, re

MODEL_ID = "dopamineaddict/qwen3-0.6b-chess-sft-v1"
tok = AutoTokenizer.from_pretrained(MODEL_ID, use_fast=True)
model = AutoModelForCausalLM.from_pretrained(MODEL_ID, torch_dtype=torch.float16, device_map="auto")
model.eval()

prompt = """Position (FEN): <YOUR_FEN>
Side to move: White
Legal moves: e2e4, g1f3, ...

Return exactly:
<uci_move>...</uci_move>
<rationale>...</rationale>
"""

inputs = tok(prompt, return_tensors="pt").to(model.device)
with torch.no_grad():
out = model.generate(**inputs, max_new_tokens=64, do_sample=False, temperature=0.0)

text = tok.decode(out, skip_special_tokens=False)
m = re.search(r"<uci_move>\s*([^<\s]+)\s*</uci_move>", text)
print("model_output:", text)
print("uci_move:", m.group(1) if m else None)
```

## Training data

- Primary dataset: `Lichess/chess-puzzles`.
- Construction approach:
  - The puzzle FEN was converted into a “player-to-move” position by applying the first UCI move from the puzzle move list.
  - The supervised label was the next UCI move in the puzzle solution sequence.
- Additional fields stored during data prep (varies by run): legal moves list, puzzle metadata, and optional Stockfish annotations.

## Training procedure

- Method: supervised fine-tuning (SFT) with LoRA adapters.
- Trainer: TRL `SFTTrainer`.
- Objective: maximize format compliance and legality (always emit a legal UCI move wrapped in `<uci_move>` tags).
- Hardware: single NVIDIA T4 GPU (typical for initial iteration).

## Evaluation

This model was primarily validated via:
- Format compliance checks (presence of `<uci_move>` and parsable UCI).
- Legality checks (output move exists in the provided legal moves list for that position).
- Quick qualitative checks on held-out puzzle positions.

No Elo or engine-match benchmark is reported for this baseline version.

## Limitations and biases

- The model is trained mostly on *tactical* puzzle positions and may be weaker in quiet/opening positions.
- The `<rationale>` field may be repetitive or generic in this version; it is not optimized for explanation quality.
- This is not a substitute for a full chess engine and can still choose suboptimal moves.

## License and attribution

- Base model: `Qwen/Qwen3-0.6B` (see upstream license and terms).
- Puzzle data: derived from the Lichess puzzles dataset.