README.md

metadata

license: apache-2.0
base_model: llava-hf/llava-1.5-7b-hf
tags:
  - machine-unlearning
  - multimodal
  - llava
  - mllmu-bench
  - remem

Ours — LLaVA-1.5-7B unlearning checkpoints

Unlearned checkpoints for MLLMU-Bench and ReMem (LLaVA-1.5-7B).

Only the vision encoder is unlearned: each trainable_state.pt holds vision-tower weights only. Load the memorized (vanilla) model and overwrite its vision-tower weights with them.

vanilla_remem_lora/                     # memorized LoRA adapter for ReMem
ours/mllmu_forget5/trainable_state.pt   # MLLMU-Bench,  5% forget
ours/mllmu_forget10/trainable_state.pt  # MLLMU-Bench, 10% forget
ours/remem_forget1/trainable_state.pt   # ReMem,  5% forget
ours/remem_forget2/trainable_state.pt   # ReMem, 10% forget

• MLLMU-Bench vanilla: yejinkim/mllmu-vanilla
• ReMem vanilla: vanilla_remem_lora/ in this repo, merged into llava-hf/llava-1.5-7b-hf