search

chingyaoc / debias_vl

Code for Debiasing Vision-Language Models via Biased Prompts
MIT License
53 stars 4 forks source link

Code for replicating FairFace experiments #1

Open ryan-caesar-ramos opened 9 months ago

ryan-caesar-ramos commented 9 months ago

Firstly, thank you for providing code for the rest of your experiments!

May I ask if there are any plans to release code for replicating the FairFace results of the paper? If not, could I at least ask what were the spurious prompts and/or positive pairs used for Section 5.2? If I'm not mistaken, Sections 4.2 and 4.3 indicate that the method is reliant on spurious prompts and positive pairs (though it seems that debiasing generative models only requires positive pairs?); Table 9 shows what these are for the classification tasks, and Table 11 shows what I assume are the class prompts for MaxSkew tasks, but which other text embeddings are required for the MaxSkew tasks and what were the corresponding prompts/prompts pairs?

Thank you!

haoyusimon commented 7 months ago

Same request here. Will really appreciate if the author could release the prompts that were used to debias CLIP for the FairFace experiment. Thank you!