README.md

metadata

dataset_info:
  features:
    - name: jpg
      dtype: image
    - name: txt
      dtype: string
    - name: njson
      dtype: string
    - name: samlens.npy
      dtype: binary
    - name: samcat.npy
      dtype: binary
  splits:
    - name: train
      num_examples: 10968539
configs:
  - config_name: default
    data_files:
      - split: train
        path: cc12m-train-*.tar
license: cc-by-4.0
task_categories:
  - zero-shot-image-classification
  - image-to-text
  - text-to-image
tags:
  - clip
  - webdataset
  - sam
  - region-phrase-alignment
size_categories:
  - 10M<n<100M

CC12M with SAM Regions and Parse-Tree Phrases

Pre-processed CC12M dataset for training PowerCLIP.

Each sample contains the original image and caption plus two precomputed annotations:

• Parse-tree phrases (.njson) — NP/PP/VP/S constituent phrases extracted via spaCy, with token indices aligned to OpenCLIP's SimpleTokenizer (CSR format).
• SAM regions (.samlens.npy + .samcat.npy) — Segment Anything Model (SAM ViT-H) region bounding boxes converted to ViT patch-grid token indices (CSR format, patch size 16, image size 224).

Format

WebDataset tar archives (2176 shards). Each sample contains:

{key}.jpg           # Image
{key}.txt           # Caption
{key}.json          # Metadata (original CC12M fields)
{key}.njson         # Parse-tree phrase indices (CSR: lengths + token IDs)
{key}.samlens.npy   # SAM region lengths array
{key}.samcat.npy    # SAM region token indices (concatenated)

Usage

import webdataset as wds

dataset = wds.WebDataset("cc12m-train-{0000..2175}.tar")
for sample in dataset:
    image = sample["jpg"]          # raw JPEG bytes
    caption = sample["txt"]        # caption string
    # SAM regions and parse-tree phrases are loaded automatically
    # by PowerCLIP's data pipeline

Or use with PowerCLIP directly:

torchrun --nproc_per_node 8 -m training.main \
    --train-data "cc12m-train-{0000..2175}.tar" \
    ...

Source

• Images & captions: Conceptual 12M (CC-BY-4.0)
• SAM regions: Segment Anything (ViT-H)
• Parse-tree phrases: spaCy en_core_web_sm