laion-coco-5m-arabic-labels

An Arabic-labelled, image-bearing derivative of CaptionEmporium/laion-coco-13m-molmo-d-7b.

The first 5,000,000 rows of the source, with the oi_yolo_pred object-detection labels translated to Arabic and the images downloaded at native resolution, republished as WebDataset.

What is Arabic here

Labels only. Every detection in oi_yolo_pred_ar carries a name_ar key. The captions (caption, caption_molmo_short/medium/long) remain English — translating them is a separate ~15M-string job and is out of scope for this release.

Method

oi_yolo_pred is the output of Ultralytics yolov8x-oiv7, whose label set is the closed Open Images V7 vocabulary of 601 class names. Across all 5M rows only 557 distinct English strings occur. So this is not a 5M-row machine-translation job:

1. the 557 strings were translated once into a reviewed glossary (glossary/oiv7_en_ar.json, also in this repo), and
2. applied by dictionary lookup.

The glossary was built from three signals — an LLM with object-detection context, a local Helsinki-NLP/opus-mt-tc-big-en-ar pass as a deliberately different paradigm, and a blind opus-mt-ar-en back-translation — then frequency-weighted triage surfaced 51 items (21% of all detections) for human review. Distinct classes were checked for collapsing to a single Arabic term; four collisions were found and resolved (Candy/Dessert, Closet/Wardrobe, Drawer/Stairs, Jug/Pitcher).

Geometry, and how to check a box is valid

Images are stored at native resolution (--resize_mode no) and never resized, so no box was rescaled by this pipeline. name and box are copied through byte-for-byte from oi_yolo_pred; only name_ar is added.

There is a caveat that matters if you use the boxes. The YOLO boxes were computed against the image as LAION saw it in 2021. Some CDNs now serve a different rendition at the same URL, so for a minority of samples the downloaded pixels no longer match that frame:

	share of samples	boxes
dimensions match LAION exactly	97.53%	valid as-is
same aspect ratio, different size	1.45%	recoverable — multiply by width / laion_original_width
different aspect ratio	1.03%	invalid — a different crop or image is being served

(1.09% of all detections sit on different-aspect samples. Measured over 28,088 sampled images.)

Every sample carries both frames, so this is fully detectable per-sample — width/height are img2dataset's own measurements of the bytes in this repo, and laion_original_width/ laion_original_height are LAION's 2021 values.

import json
m = r["json"]
if (m["width"], m["height"]) == (m["laion_original_width"], m["laion_original_height"]):
    boxes_ok = True                     # ~98% of samples
else:
    s = m["width"] / m["laion_original_width"]
    boxes_ok = abs(m["width"]/m["height"] - m["laion_original_width"]/m["laion_original_height"]) < 0.02
    # if boxes_ok, multiply every box coordinate by s

Nothing here is silently wrong; the images and captions are unaffected either way.

Format

WebDataset. Each sample is {key}.webp (native resolution, WebP quality 85), {key}.txt (the LAION caption) and {key}.json with all source columns.

from datasets import load_dataset
ds = load_dataset("oddadmix/laion-coco-5m-arabic-labels", streaming=True, split="train")
r = next(iter(ds))
import json; print(json.loads(r["json"]["oi_yolo_pred_ar"])[0])
# {"name": "House", "class": 257, "confidence": 0.90153, "box": {...}, "name_ar": "منزل"}

Column notes:

• img2dataset reserves key, width, height, original_width, original_height, status, error_message, exif, sha256 — LAION's own values are preserved under laion_*. The unprefixed names are img2dataset's own freshly-measured values.
• all_captions and all_similarities are JSON-encoded strings; use json.loads().

Coverage

rows in the 5M slice	5,000,000
detections translated	20,005,352
distinct labels	557
shards published	798 (data/shard_00 … data/shard_06)
rows attempted in those shards	1,594,185
images published	1,117,727
download success rate	70.1%
bytes of image data	~132 GB
images with exactly-matching dimensions	97.53%

Counts are the sum of the per-shard *_stats.json files in this repo, so they are recomputable from it. Roughly 30% of LAION URLs are dead, which is why the image count is well below the row count. This is expected and unavoidable for a URL-only source, and the shard set is a prefix of the 5M slice rather than a sample of it.

What was trained on this

oddadmix/Nawah-VL-50M-detect-ar-stage3, a 145M Arabic dense-object-detection VLM, was trained on the oi_yolo_pred_ar column of this dataset. Its recipe, which is one reasonable default for using these boxes as supervision:

• keep detections at confidence >= 0.40, cap at 10 per image;
• sort the survivors largest-area-first (source order is confidence-descending, which is very nearly uncorrelated with area and therefore not a learnable emission order);
• drop rows where nothing survives — that is 23.7% of images;
• cap images at 512 px on the long side and drop aspect ratios worse than 5:1;
• hold out whole shards, not rows, so the eval set cannot share an image with training.

That leaves 838,863 usable rows over the 794 shards that existed at the time, at 2.90 objects per row. The training corpus builder and the evaluator are in code/ of that model repo.

Honest caveats

• The labels are machine-generated. They are one YOLO checkpoint's opinion, not human annotation. A missing object is far more often a detector miss than a genuine absence, so this is supervision for detection-style behaviour rather than a ground-truth detection benchmark.
• The vocabulary is closed. 601 Open Images V7 classes, of which only 557 ever occur. Anything outside it is unlabelled, in any language.
• The Arabic is glossary Arabic. 557 strings translated once and applied by lookup, with 51 frequency-weighted items human-reviewed. It is consistent by construction, but a term that is wrong is wrong everywhere it appears.
• ~1% of samples have unusable boxes, and another ~1.5% need rescaling — see the geometry section above, and check per sample rather than assuming.
• The captions are still English. Only the detection labels are Arabic.

Provenance and takedown

url and laion_sha256 are retained per sample. Re-hosting image bytes is a deliberate change of posture from LAION, which ships URLs only. To request removal of an image, open a discussion on this repo with the url or laion_sha256.

License and attribution

cc-by-sa-4.0, inherited from the source. Attribution required to Caption Emporium (laion-coco-13m-molmo-d-7b) and LAION. Object labels originate from Open Images V7 via Ultralytics yolov8x-oiv7.