search

UCSC-VLAA / Recap-DataComp-1B

This is the official repository of our paper "What If We Recaption Billions of Web Images with LLaMA-3 ?"
https://www.haqtu.me/Recap-Datacomp-1B/
115 stars 1 forks source link

Have u compare tthe Recap DataComp dataset trained old llava old only? #3

Open MonolithFoundation opened 3 months ago

MonolithFoundation commented 3 months ago

As far as I konw, the llava with llama3 8b newest version already get a very good result without RecapDataComp1B

Just wondering, how does the dataset contribute to the performance without changing the model.

ImKeTT commented 3 months ago

Thank you for your interest in our work! We didn't fine-tune the LLaMA3-powered LLaVA using our RecapDataComp1B. Instead, we use the powerful LLaMA3-powered LLaVA to recaption the DataComp-1B and the resulting dataset is our RecapDataComp1B.

MonolithFoundation commented 3 months ago

does the dataset only can boost lava original performance?

---- Replied Message ---- | From | Haoqin @.> | | Date | 06/15/2024 11:22 | | To | @.> | | Cc | @.>@.> | | Subject | Re: [UCSC-VLAA/Recap-DataComp-1B] Have u compare tthe Recap DataComp dataset trained old llava old only? (Issue #3) |

Thank you for your interest in our work! We didn't fine-tune the LLaMA3-powered LLaVA using our RecapDataComp1B. Instead, we use the powerful LLaMA3-powered LLaVA to recaption the DataComp-1B and the resulting dataset is our RecapDataComp1B.

— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: @.***>

ImKeTT commented 3 months ago

Sorry, but I'm a little confused, what do you mean by 'boosting LLaVA original performance'? As we didn't use our RecapDataComp1B to fine-tune a LLaVA model.

MonolithFoundation commented 3 months ago

Just want to make sure, is the sysnthsis data can boost performance or not. You should keep model same to compare it. using a new LLM then you should compare same llava-llama3 model which offcially released.