HIPPO-Video Simulation Framework

This repository contains the official code for generating YouTube watch histories using an LLM-based user simulator.

📖 Proposed in the COLM 2025 paper:
HIPPO-VIDEO: Simulating Watch Histories with Large Language Models for History-Driven Video Highlighting

You can access the full dataset generated by this simulation pipeline on Hugging Face.

---

🚀 Getting Started

1. Clone and set up environment

git clone https://github.com/jeongeunnn-e/HIPPO-Video.git
cd HIPPO-Video

# Conda (recommended)
conda create -n hippo python=3.10 -y
conda activate hippo
pip install -r requirements.txt

⸻

2. Prepare config and seed data

You need to provide a configuration and input data in JSON format. We include an example seed file: seed_data.json.

Example: config.json

{
  "data_path": "your_path/seed_data.json",
  "save_path": "your_path/outputs/",
  "donwload_path": "your_path/downloads/",
  "model_name": "gpt-4o",
  "max_length": 10,
  "OPENAI_API_KEY": "your_openai_key"
}

📑 Example: seed_data.json

[
  {
    "topic": "Clothes",
    "sub_topic": "Shoes",
    "feature": "informative",
    "initial_query": "how shoes are made from start to finish"
  },
  {
    "topic": "Music",
    "sub_topic": "Jazz",
    "feature": "emotional",
    "initial_query": "best emotional jazz solos"
  }
]

You can include multiple seeds to generate multiple simulated sessions.

⸻

3. Run simulation

python run.py