naver-ai/M3T
收藏资源简介:
--- license: cc-by-4.0 task_categories: - feature-extraction language: - en tags: - multimodal - M3T size_categories: - 100M<n<1B --- # MuCo: Multi-turn Contrastive Learning for Multimodal Embedding Model (CVPR 2026) [[Paper]](https://arxiv.org/abs/2602.06393) [[GitHub]](https://github.com/naver-ai/muco) ## Dataset Overview This repository contains the training data used in **MuCo**. ``` . ├── M3T_train/ # M3T training data (meta only, images must be downloaded separately) │ └── M3T.parquet ├── MMEB_train/ # MMEB meta files with generated captions │ ├── A-OKVQA/ │ ├── CIRR/ │ ├── MSCOCO/ │ ├── ... │ └── WebQA/ ├── download_M3T_images.py └── README.md ``` ### M3T M3T is a multi-turn multimodal training dataset. The meta file (`M3T_train/M3T.parquet`) contains query text, image paths, positive text pairs, and image URLs for downloading. Since image files are not included in this repository due to their size, you need to download them using the provided script. ### MMEB `MMEB_train/` contains meta files for each MMEB benchmark subset. Each subfolder includes a `*_generated_sentence.parquet` file, which stores the generated captions for images as described in the paper. ## Getting Started ### 1. Download the dataset from Hugging Face ```bash pip install huggingface_hub hf download naver-ai/M3T --repo-type dataset --local-dir M3T cd M3T ``` ### 2. Download M3T images The images are downloaded from their original source URLs using the provided script. It runs with all available CPU workers and retries failed downloads up to 5 times with exponential backoff. ```bash python download_M3T_images.py ``` After downloading, successfully fetched entries are saved to `M3T_train/M3T_success.parquet` to maintain meta-image consistency. Use this file for training instead of `M3T.parquet` if any downloads fail. ## License We distribute the image URL-text samples and metadata under [Creative Commons CC-BY-4.0](https://creativecommons.org/licenses/by/4.0/) license. The individual images are sourced from [DataComp-1B](https://github.com/mlfoundations/datacomp) and are under their own copyrights.




