search

pipilurj / bootstrapped-preference-optimization-BPO

code for "Strengthening Multimodal Large Language Model with Bootstrapped Preference Optimization"
Apache License 2.0
45 stars 1 forks source link

The datasets with negative responses. #4

Closed darkpromise98 closed 3 months ago

darkpromise98 commented 3 months ago

Thanks for your awesome work.

As shown in the Tab. 2, the preference training datasets have 150K samples from three data sources (ShareGPT-V, LLaVAR, LLaVA-Instruct). Do you have plans to release all ready-made preference datasets with negative responses (including Image-Weakened prompting and Error Injection)? I think the ready-made datasets can help the community and researchers follow your work better.

pipilurj commented 3 months ago

Thank you very much for your interest. We have uploaded the negative responses on huggingface in this link:https://huggingface.co/datasets/renjiepi/BPO