12 Million Image-Text Pairs for Vision-Language Pre-training

Conceptual 12M contains 12 million image-text pairs intended for vision-and-language pre-training. It was created by Google Research using a relaxed version of the data collection pipeline from Conceptual Captions 3M.

Use Cases

• Train a vision-language model on 12 million image-text pairs for tasks like image captioning.
• Fine-tune a contrastive learning model using the large-scale image-text pairs for cross-modal retrieval.
• Use the image-text pairs for pre-training multimodal transformers to improve zero-shot classification.

Strengths

• Contains 12 million image-text pairs, providing a large-scale resource for pre-training.
• Built by Google Research, indicating a systematic collection and curation process.
• Based on a pipeline from the established Conceptual Captions 3M dataset.

Limitations

• Specific column structure, image sources, and annotation quality details are unknown.
• The dataset's geographic and temporal coverage is not specified, which may limit generalizability.
• The 'relaxed' collection pipeline may introduce more noise compared to its predecessor.

Provenance

Source

google-research-datasets

Collection Method

Collected using a relaxed version of the pipeline from Conceptual Captions 3M.

Freshness

Last updated on 2024-01-18.

Geography

Region tag indicates 'us', but specific spatial coverage is unknown.