TiC-DataComp is a dataset for the Time-Continual Learning (TiC) benchmark for vision-language models, released by Apple in 2024. The associated paper is "TiC-CLIP: Continual Training of CLIP Models", and the dataset aims to address the high cost of continuously updating foundation models by providing large-scale time-series data to evaluate and improve the temporal robustness of models.

The dataset contains over 12.7 billion timestamped image-text pairs spanning from 2014 to 2022, sourced from CommonCrawl. Researchers can use this dataset to design and evaluate time-continual learning methods for large-scale image-text models, measure the temporal robustness of existing models through dynamic evaluation, and validate the effectiveness of replay-based training strategies in reducing computational resource consumption.

Dataset Composition

The dataset primarily includes the following structures for training and evaluation at different temporal granularities:

• Monthly training data: tic-datacomp_training_monthly/.npy, containing a list of unique identifiers (UIDs) for each month.
• Annual training data: tic-datacomp_training_yearly_noeval/.npy, containing lists of UIDs for each year with the annual evaluation sets removed.
• Retrieval evaluation data: tic-datacomp_retrieval_evals_year2uids, containing UIDs for each year in the TiC-DataComp-Retrieval evaluation tasks.
• Net evaluation data: tic-datacompnet_year2uids, containing UIDs for each year in the TiC-DataCompNet evaluation tasks.

Citation

@inproceedings{garg2024tic,
  title={TiC-CLIP: Continual Training of CLIP Models},
  author={Garg, Saurabh and Farajtabar, Mehrdad and Pouransari, Hadi and Vemulapalli, Raviteja and Mehta, Sachin and Tuzel, Oncel and Shankar, Vaishaal and Faghri, Fartash},
  booktitle={The Twelfth International Conference on Learning Representations (ICLR)},
  year={2024},
  url={https://openreview.net/forum?id=TLADT8Wrhn}
}