ljcnju/Qwen3-8B-Chess-SFT
The ljcnju/Qwen3-8B-Chess-SFT model is an 8 billion parameter language model based on the Qwen3 architecture, fine-tuned specifically for chess-related tasks. This model is the second stage of supervised fine-tuning (SFT) as described in the paper arXiv:2509.24239. It leverages a 32768 token context length and is optimized for understanding and generating content within the domain of chess, making it suitable for applications requiring specialized chess knowledge.
Loading preview...
Model Overview
The ljcnju/Qwen3-8B-Chess-SFT is an 8 billion parameter model built on the Qwen3 architecture, specifically designed for chess-related applications. This model represents the second stage of supervised fine-tuning (SFT) as detailed in the research paper arXiv:2509.24239.
Key Capabilities
- Chess Domain Specialization: Fine-tuned on the ChessArena_Training_Dataset, indicating a strong focus on chess-specific language and patterns.
- Qwen3 Architecture: Benefits from the underlying capabilities of the Qwen3 base model.
- Extended Context Window: Features a 32768 token context length, allowing for processing longer chess game sequences or discussions.
Training Details
The model was trained with a learning rate of 5e-06 over 3 epochs, utilizing a total batch size of 32 across 8 GPUs. The training process showed a consistent reduction in both training and validation loss, with a final validation loss of 0.4887. Key hyperparameters included a cosine learning rate scheduler with a 0.1 warmup ratio and the AdamW_TORCH optimizer.