Confusion about the LlavaGuard versions
Hi, I am quite confused on the current situation with the LlavaGuard models that have been released.
In the first version of the paper, you have released:
7B, 13B, 34B models
Then in the second version, you talked only about a 0.5B and 7B version of the models and you do not mention the 13 and 34 versions.
In the website, the reported benchmarks are extracted from the first version of the paper.
Is this the 7B version mentioned in the second version of the paper? And the same for the 0.5B version?
Hi, thanks for your question and sorry for the confusion. We will be updating the website soon to reflect the latest information.
To clarify, there are generally three versions of LlavaGuard models released:
Initial Version: This includes the 7B, 13B, and 34B models based on the original Llava models of the same sizes.
Intermediate Version (v1.1): This update still uses the 7B and 13B Llava models but incorporates an improved dataset with enhanced "guided rationales," as detailed in the second version of our paper.
Newest Version (v1.2): This version is based on the Llava-OV models and includes the 0.5B and 7B sizes. These models use the latest dataset and demonstrate the best overall performance.
Regarding the benchmarks on the website, they currently reflect the initial version’s results. The 7B model you see mentioned in the second version of the paper corresponds to the latest v1.2 release.
I hope this clears things up! Let me know if you have any more questions.
Thank you very much!
We also reordered the space a bit and adjusted the names, now it should be even clearer. Thanks for the heads-up!