README.md

metadata

dataset_info:
  - config_name: default
    features:
      - name: image
        dtype: image
      - name: keywords
        dtype: string
    splits:
      - name: train
        num_bytes: 2045209850.972
        num_examples: 24996
    download_size: 1935601893
    dataset_size: 2045209850.972
  - config_name: embeddings_clip-ViT-B-32
    features:
      - name: embeddings_clip-ViT-B-32
        sequence: float32
    splits:
      - name: train
        num_bytes: 51291792
        num_examples: 24996
    download_size: 66797740
    dataset_size: 51291792
  - config_name: embeddings_metaclip-2-worldwide-s16-384
    features:
      - name: embeddings_metaclip-2-worldwide-s16-384
        sequence: float32
    splits:
      - name: train
        num_bytes: 38493840
        num_examples: 24996
    download_size: 53998486
    dataset_size: 38493840
  - config_name: embeddings_metaclip-2-worldwide-s16-384-eng-32768
    features:
      - name: embeddings_metaclip-2-worldwide-s16-384-eng-32768
        sequence: float32
    splits:
      - name: train
        num_bytes: 38493840
        num_examples: 24996
    download_size: 53998780
    dataset_size: 38493840
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
  - config_name: embeddings_clip-ViT-B-32
    data_files:
      - split: train
        path: embeddings_clip-ViT-B-32/train-*
  - config_name: embeddings_metaclip-2-worldwide-s16-384
    data_files:
      - split: train
        path: embeddings_metaclip-2-worldwide-s16-384/train-*
  - config_name: embeddings_metaclip-2-worldwide-s16-384-eng-32768
    data_files:
      - split: train
        path: embeddings_metaclip-2-worldwide-s16-384-eng-32768/train-*

Unsplash Lite

Unsplash Lite Dataset.
This dataset contains:

• a subset default with about 25k images and related keywords when available. Keywords are eparated by ; and note that we kept only those where the confident score indicated by Unsplash is higher than 90%
• a subset embeddings_clip-ViT-B-32 which contains precomputed embeddings of the images via the clip-ViT-B-32 model by OpenAI
• a subset embeddings_metaclip-2-worldwide-s16-384 which contains precomputed embeddings of the images via the metaclip-2-worldwide-s16-384 model by Meta
• a subset embeddings_metaclip-2-worldwide-s16-384-eng-32768 which contains precomputed embeddings of the images via the metaclip-2-worldwide-s16-384-eng-32768 model by AlphaEdge

These three precomputed embeddings subsets are useful for tutorial notebooks in the Sentence Transformers documentation showing how to use the lib to perform (multi-lingual) 0-shot image classification, monolingual/multilingual, image clustering and image de-duplication.

License

https://unsplash.com/license

TL;DR:
Unsplash images are designed to be used freely, and our license reflects this.

• It is possible to download and use all the images for free
• For commercial and non-commercial purposes
• No permission required (an assignment will always be appreciated!)

What is not allowed 👎

• It is not allowed to sell the images without significant modification.
• Compiling images from Unsplash to replicate a similar or competing service.

Credit to lbourdois for collecting and uploading this dataset.