DataComp-1B Image-Text Pair Metadata

## Overview DataComp-1B is the metadata index for ~1.4 billion image-text pairs selected from the 12.8B-sample CommonPool via the DataComp filtering competition. It is the training corpus behind several state-of-the-art open CLIP checkpoints. This listing exposes the parquet metadata shards (image URLs, alt-text captions, and precomputed CLIP features / similarity scores) — not the images themselves, which remain under their original copyrights and must be fetched by the consumer. Format: parquet; modality: tabular metadata referencing image+text. ## Schema - `url` — string — source URL of the image - `text` — string — associated alt-text / caption - `uid` — string — unique sample identifier - `original_width` — int — original image width in pixels - `original_height` — int — original image height in pixels - `sha256` — string — image content hash for verification - `clip_b32_similarity_score` — float — CLIP ViT-B/32 image-text cosine similarity - `clip_l14_similarity_score` — float — CLIP ViT-L/14 image-text cosine similarity - `face_bboxes` — array — detected face bounding boxes (for blur/filter pipelines) - +a few more metadata columns per the official schema ## Sources - HuggingFace: https://huggingface.co/datasets/mlfoundations/datacomp_1b — license: CC-BY-4.0 - Project site & code: https://www.datacomp.ai/ and https://github.com/mlfoundations/datacomp ## Methodology The upstream CommonPool was crawled from Common Crawl WARC files between 2014–2022, with image-text pairs extracted from HTML alt attributes. The DataComp-1B subset was selected by combining (a) image-text CLIP similarity filtering and (b) ImageNet-based image clustering, as described in Gadre et al., "DataComp: In search of the next generation of multimodal datasets" (NeurIPS 2023). Precomputed CLIP B/32 and L/14 embeddings are distributed to allow re-filtering without re-encoding. NSFW detection and face-bbox annotations are included so downstream users can apply their own safety filters. ## Known gaps & limitations - URLs decay over time; expected fetch success has dropped materially since 2023 release and continues to decline. - Heavy English / Western web bias; long-tail languages underrepresented. - Captions are alt-text, which is often noisy, SEO spam, or non-descriptive. - Source documents NSFW and face detections but does not guarantee removal — downstream filtering is the consumer's responsibility. - Individual images remain under original copyrights; CC-BY-4.0 covers only the URL/text/metadata table. ## Intended use & out-of-scope - Intended: training/fine-tuning CLIP-family multimodal models, ablation studies on data curation, retrieval research, multimodal eval set construction. - Out-of-scope: redistribution of the underlying images (not licensed here); use cases requiring guaranteed-available image URLs; deployments without independent NSFW/PII filtering. _Federated dataset: 2,664 parquet shards, 316.53 GB total. Queries and downloads stream through the DataBazaar API._ _PII signals: us_phone×4, credit_card_candidate×2 (0.3≤score<0.7) present in the sample. Common in public datasets (papers, logs) but worth knowing before joining with private data._ Original supplier listing: DataComp-1B: Image-Text Pair Metadata (1.4B Samples) Metadata (URLs, captions, CLIP features) for ~1.4B image-text pairs from DataComp-1B, the curated subset of CommonPool used to train state-of-the-art CLIP models. CC-BY-4.0, parquet format.