Upload Llama Unlearning Checkpoints

[shariqahn]

I would like to use the Llama model for evaluation. I see you put the metrics for models finetuned on various splits of the data in the data directory. I also see you uploaded the model finetuned on the entire dataset for Llama, but not ones with unlearning performed on it like you did for Phi. It would be great to have access to those for baselines since I am running into some architecture limitations for running the forget script.

[zhilif]

will do so soon!

[shariqahn]

Thank you! I have a deadline coming up. If there is any way to upload these soon, I would really appreciate. By the new year would be great if at all possible :)

[zhilif]

@shariqahn I will try my best. Uploading every checkpoint for every unlearning method might take a while. Is there a setting that you want us to prioritize?

[shariqahn]

Preference is the forget10 setting with whatever models are quickest to upload. Appreciate it!

[shariqahn]

Hi! Do you happen to have a sense of when you will able to upload things? Would just be helpful to have an idea to plan out things for my deadline.

[zhilif]

Hi! Do you happen to have a sense of when you will able to upload things? Would just be helpful to have an idea to plan out things for my deadline.

Will do by Sunday

[shariqahn]

Thank you!

[zhilif]

@shariqahn The uploading script is still running but some checkpoints are there. For example:
https://huggingface.co/locuslab/llama2-7b_KL_1e-05_forget10/tree/checkpoint-12.

You can pick checkpoints in their corresponding git branch