Implementation instructions and details for MMSep
Follow your desired MLLM's repository for installation.
Copy and paste functions in mm_separators.py to your desired MLLM's Language model class and call in layered transformer loop. Copy and paste cache.py to your language model codebase, and replace the original Cache class with MMSepCache.
Collect answers in LLaVA-v1.5 validation jsonl manner and use auto_eval.py with your own evaluation model api.
More details are still under preperation.