36 lines
1.1 KiB
Markdown
36 lines
1.1 KiB
Markdown
|
---
|
|||
|
|
library_name: transformers
|
|||
|
|
tags: []
|
|||
|
|
---
|
|||
|
|
|
|||
|
|
# Model Card for chess-qwen-finetuned-v2
|
|||
|
|
Fine-tuned Qwen2.5-0.5B-Instruct for chess move prediction.
|
|||
|
|
Given a board position in FEN notation and a list of legal moves, the model outputs the best move in UCI format.
|
|||
|
|
|
|||
|
|
## Model Description
|
|||
|
|
- **Developed by:** Evelien van Driel
|
|||
|
|
- **Model type:** Causal Language Model (decoder-only)
|
|||
|
|
- **Language(s) (NLP):** English
|
|||
|
|
- **Finetuned from model:** Qwen/Qwen2.5-0.5B-Instruct
|
|||
|
|
|
|||
|
|
## Direct Use
|
|||
|
|
Chess move prediction as part of INFOMTALC 2026 (Utrecht University).
|
|||
|
|
Used inside a TransformerPlayer class that queries the model given a FEN position.
|
|||
|
|
|
|||
|
|
## Training Details
|
|||
|
|
|
|||
|
|
### Training Data
|
|||
|
|
aicrowd/ChessExplained dataset, examples 0–100,000 (100k positions).
|
|||
|
|
First fine-tuned v1 on examples 0–50,000 (chess-qwen-finetuned), then continued fine-tuning from v1 on examples 50,000–100,000 (v2).
|
|||
|
|
Moves are Stockfish-approved.
|
|||
|
|
|
|||
|
|
### Training Hyperparameters
|
|||
|
|
|
|||
|
|
- **Training regime:**
|
|||
|
|
- Base model: Qwen/Qwen2.5-0.5B-Instruct
|
|||
|
|
- Method: LoRA
|
|||
|
|
- Epochs: 3
|
|||
|
|
- Batch size: 16
|
|||
|
|
- Learning rate: 2e-4
|
|||
|
|
- Hardware: Google Colab (T4 GPU)
|