Dataset Viewer
Auto-converted to Parquet Duplicate
num_input
int64
num_after_dedup
int64
num_after_clip
int64
num_sharded
int64
exact_dups
int64
near_dups
int64
dedup_rate
float64
num_shards
int64
num_workers
int64
hash_kind
string
elapsed_sec
float64
images_per_sec
float64
band_pairs_checked
int64
clip_ran
bool
extra
dict
3,840
3,000
2,536
2,536
637
203
0.21875
3
8
dhash
14.826966
258.987579
556,812
true
{ "clip_threshold": 0.1, "clip_score_min": 0.036224767565727234, "clip_score_max": 0.3183661103248596 }

parallel-image-text-dataset-builder (sample)

A small representative sample from the parallel-image-text-dataset-builder pipeline: it ingests image-text pairs, removes near-duplicates with perceptual-hash (dhash) LSH-style bucketing, filters weak pairs by CLIP image-text similarity, and writes fixed-size WebDataset-style tar shards.

Contents

  • shard-00002.tar - one WebDataset-style shard (536 samples). Each sample is two members sharing a key: {key}.jpg (image) and {key}.txt (caption).
  • stats.json - machine-readable statistics from the full run that produced this sample.

Generation method

Fully synthetic. Base images are procedural RGB patterns built from a sum of random low-frequency plane waves plus a bright blob and a per-image color tint, so each base image has a distinctive perceptual hash while a planted resize + JPEG-recompress copy hashes close to it. Captions are templated (a {color} {shape} over a {scene}) and only loosely tied to the abstract images, so absolute CLIP scores are low by construction.

The published shard is the output of the full pipeline:

  1. Perceptual hashing (dhash, 9x8 grid, 64-bit) over all inputs.
  2. Near-duplicate removal via banded LSH bucketing + union-find (5-bit hamming threshold, 8 bands).
  3. CLIP image-text similarity filtering (open_clip ViT-B-32 laion2b_s34b_b79k) at score threshold 0.10.
  4. Fixed-size tar sharding (1000 samples/shard).

Run this sample came from

Measured on a single NVIDIA RTX 5090 (24 GB). Input: 3840 synthetic pairs (3000 unique base images + 840 planted near-duplicates).

stage value
input pairs 3840
after dedup 3000 (removed 840: 637 exact, 203 near)
dedup rate 0.2188 (equals the planted duplicate fraction)
after CLIP filter (threshold 0.10) 2536
shards written 3 (this repo ships shard 2, 536 samples)
band pairs checked 556,812 vs 7,370,880 all-pairs

This is a small-scale run that exercises the full path end to end; the numbers are real measurements, not a large-corpus benchmark.

Reading a shard

import tarfile

with tarfile.open("shard-00002.tar") as tar:
    jpgs = [n for n in tar.getnames() if n.endswith(".jpg")]
    print(len(jpgs), "images")

License

MIT.

Downloads last month
12