Prompt Engineering Improves Video Understanding

This repository contains the code for our paper, "Prompt Engineering Improves Video Understanding."

Directory Structure

./llama: Contains code utilizing LLaMA 3.1 for generating accuracy scores, rewriting original questions, and generating automatic prefixes.

./llava_next_video: Includes code for using the LLaVA-NeXT-Video-7B-DPO model to perform inference across all prompt settings.

./llava_one_vision: Contains code for using the LLaVA-OneVision-Qwen2-7B-OV-Chat model to perform inference across all prompt settings.

./qwen2_vl: Features code for using the Qwen2-VL-7B-Instruct model to perform inference across all prompt settings.

Scripts and Examples

./inference.sh: An example script demonstrating how to perform inference on the TGIF dataset using the LLaVA-NeXT-Video-7B-DPO model with all prompt settings.

./prompts.py: Contains all the prompts used in our experiments.