DeepSeek R1 Checkpoint 4

This is checkpoint 4 from training the DeepSeek R1 model on mathematical reasoning tasks using GRPO.

Model: deepseek-ai/deepseek-r1-distill-qwen-7b
Checkpoint Step: 4
Training Details: Trained on 8 NVIDIA B200 GPUs with batch size 6 per GPU, 6 epochs.