xtuner

xtuner copied to clipboard

Dear XTuner Contributors,

Thank you for providing the open-source code for MLLM. I noticed that the process_hf_dataset function in the LLaVADataset takes several minutes to preprocess text data each time the program starts. In contrast, some other repositories (e.g., Original LLaVA, LLaVA-Next) handle text data preprocessing during training.

To address this, I have proposed an enhancement to the LLaVA dataset processing by introducing optional text preprocessing. This modification eliminates the need to preprocess text data at runtime, improving efficiency.

I kindly request you review my code and consider merging my PR.

Best regards,

wanghao9610