ljcnju/Qwen3-8B-Chess-SFT

Model description

This model is the SFT stage two model of paper https://arxiv.org/abs/2509.24239.

Training and evaluation data

Training dataset: https://huggingface.co/datasets/ljcnju/ChessArena_Training_Dataset

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 5e-06
• train_batch_size: 1
• eval_batch_size: 1
• seed: 42
• distributed_type: multi-GPU
• num_devices: 8
• gradient_accumulation_steps: 4
• total_train_batch_size: 32
• total_eval_batch_size: 8
• optimizer: Use OptimizerNames.ADAMW_TORCH with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: cosine
• lr_scheduler_warmup_ratio: 0.1
• num_epochs: 3.0

Training results

Training Loss	Epoch	Step	Validation Loss
0.5441	0.5667	500	0.5742
0.462	1.1326	1000	0.5241
0.4573	1.6993	1500	0.4992
0.4061	2.2652	2000	0.4926
0.3979	2.8320	2500	0.4887

Framework versions

• Transformers 4.52.4
• Pytorch 2.8.0+cu128
• Datasets 3.6.0
• Tokenizers 0.21.1