Video-LLaVA-7B-hf-CinePile

Model Card for Video-LLaVA - CinePile fine tune

Video multimodal research often emphasizes activity recognition and object-centered tasks, such as determining "what is the person wearing a red hat doing?" However, this focus overlooks areas like theme exploration, narrative and plot analysis, and character and relationship dynamics. CinePile addresses these areas in their benchmark and they find that Large Language Models significantly lag behind human performance in these tasks. Additionally, there is a notable disparity in performance between open and closed models.

In our initial fine-tuning, our goal was to assess how well open models can approach the performance of closed models. By fine-tuning Video LlaVa, we achieved performance levels comparable to those of Claude 3 (Opus).

Results

Fine-tuned model taking as starting point Video-LlaVA to evaluate its performance on CinePile.

Model Sources

<!-- Provide the basic links for the model. -->

• Repository: Github with fine-tunning and inference notebook.

Uses

Although the model can answer questions based on the content, it is specifically optimized for addressing CinePile-related queries. When the questions do not follow a CinePile-specific prompt, the inference section of the notebook is designed to refine and clean up the text produced by the model.