MedGenVidQA 2026 Starter Kit

Please install Anaconda to create a conda environment as follows:

# preparing environment
conda create -n medgenvidqa2026 python=3.10
conda activate medgenvidqa2026
pip install -r requirements.txt

Install Java Dependency

wget https://github.com/adoptium/temurin21-binaries/releases/download/jdk-21.0.1+12/OpenJDK21U-jdk_x64_linux_hotspot_21.0.1_12.tar.gz
mkdir -p $HOME/jdk
tar -xzf OpenJDK21U-jdk_x64_linux_hotspot_21.0.1_12.tar.gz -C $HOME/jdk
export JAVA_HOME="$HOME/jdk/jdk-21.0.1+12"
export PATH="$JAVA_HOME/bin:$PATH"

Install Pyserini Dependency

conda install -c pytorch faiss-cpu -y

Download HowTo100M caption which contains majority of the YouTube videos subtitles

wget https://www.rocq.inria.fr/cluster-willow/amiech/howto100m/raw_caption.zip

Place the JSON file in data/HowTo100M

Download the TestDataset and VideoCorpus shared via CodaBench and place them in

data/TestDataset and data/VideoCorpus respectively.

Download Llama2

Before downloading, you need to agree to Meta's license terms by visiting here: https://huggingface.co/meta-llama/Llama-2-7b-chat-hf

You may need to fill out the form to agree to the license terms. Once your request approved, run the following:

  pip install huggingface_hub
  huggingface-cli login
  huggingface-cli download meta-llama/Llama-2-7b-chat-hf --local-dir llama-2-7b-chat-hf

Download the fine-tuned adapter from here and place them in data/FineTunedAdapter [Task B only]

Change directory to

cd src/

Download and index pubmed baseline

./build_pubmed_index.sh

It will index 28,372,706 PubMed documents.

Process and index video subtitles

./build_video_index.sh

Run Task A Baseline

 python task_a_baseline.py

The submission ready files (3 baseline approaches) will be saved in data/BaselineResults directory.

Run Task B Baseline

 python task_b_baseline.py

Run Task C Baseline

 python task_c_baseline.py