asgaardlab/VideoGameBunny-Dataset

--- license: mit task_categories: - image-to-text language: - en tags: - VideoGames pretty_name: VideoGameBunny-Dataset size_categories: - 100K<n<1M --- # VideoGameBunny Instruction Following Dataset [Paper](https://huggingface.co/papers/2407.15295) - [Website](https://videogamebunny.github.io/) ## Overview We present a comprehensive dataset of 185,259 high-resolution images from 413 video games, sourced from YouTube videos. This dataset addresses the lack of game-specific instruction-following data and aims to improve the ability of open-source models to understand and respond to video game content.  ## Dataset Composition Our dataset includes various types of instructions generated for these images using different large multimodal models: 1. Short captions 2. Long captions 3. Image-to-JSON conversions 4. Image-based question-answering pairs ## Dataset Statistics | Task | Generator | Samples | |------|-----------|---------| | Short Captions | Gemini-1.0-Pro-Vision | 70,673 | | Long Captions | GPT-4V | 70,799 | | Image-to-JSON | Gemini-1.5-Pro | 136,974 | | Question Answering | Llama-3, GPT-4o | 81,122 |