Paper: Empowering Reliable Visual-Centric Instruction Following in MLLMs
Github: KerenWLHe/VC-IFEval
š This is the official repo of VC-IFEngine datasets.
š Image sources: LLaVA-Instruct, Visual Genome, ALLaVA-4V
SFT ā messages format (role: user / assistant):
{
"messages": [
{"role": "user", "content": "
\n{instruction}"},
{"role": "assistant", "content": "{response}"}
],
"images": ["Data/CC3M/images/example.jpg"]
}
DPO āā¦ See the full description on the dataset page:
https://huggingface.co/datasets/WoofWoof/VCIF-10k.
