VideoAVE

An open-source video-to-text attribute value extraction dataset.

Dataset

• 224k training data, 25k testing data, 14 domains, 172 unique attributes

Getting Started

Data Pre-processing

For processing from raw data:

cd data_processing
python raw_data_processing.py --data_name raw_meta_All_Beauty --output_path raw_meta_All_Beauty.csv

For implementing CLIP-MoE data curation steps:

python CLIP-MoE.py

The output will be the same as training data and testing data provided in Dataset folder.

Benchmarks Inference

We support four benchmarks: Video-LLaVA, VideoLLaMA3, InternVideo2.5, and Qwen2.5-VL.

Here is an example for Qwen2.5-VL:

cd benchmarks_inference
python qwen.py

Benchmarks Finetuning

We use LLaMA-Factory for fine-tuning Qwen2.5-VL. data_info and yaml files are provided in model_training. Here is the fine-tuning steps:

• Clone LLaMa-Factory Repo and setup the environment.
• Generating training data in LLaMA-Factory data format. Please refer to:

 ./data_processing/training_data_processing.ipynb

• Update yaml file as:

 ./model_training/qwen2_5vl_full_sft.yaml

• Start training

Evaluation

 evaluation.ipynb