oddadmix/laion-coco-nllb-arabic-filtered

LAION-COCO-NLLB, Arabic slice, filtered

The arb_Arab captions of visheratin/laion-coco-nllb, extracted and filtered for training an Arabic captioning model. This is the exact corpus behind oddadmix/Nawah-VL-25M.

split	rows	kept from
train	753,883	878,978 (85.8%)
test	14,407	14,906 (96.7%)

Columns

column	type	notes
id	string	the source dataset's image id
url	string	image URL, not the image. See below.
caption	string	the arb_Arab caption, whitespace-normalised
score	float	the source dataset's score

Images are not included, only URLs. The source is a 2021 LAION crawl, so a fraction of the links are dead: fetching all 753,883 with img2dataset recovered about 79% of them. Two practical notes from doing that at scale, both in the training code linked below:

• Sort the URL list by hostname before crawling. In caption order, nearly every DNS lookup is a cache miss, and a few hundred concurrent cold lookups is enough to make a local systemd-resolved stub return SERVFAIL. 4.3 URLs per host on average, and the top 1000 hosts cover 65% of the corpus.
• Retry the transient failures. Timeouts and resolution errors are worth a second attempt; 404/403/400 are not. A retry pass recovered 139,532 of 141,414 failures.

Filtering

Starting from every row of the source dataset:

dropped for	train
duplicate caption	107,396
length outside 4-40 tokens	14,973
under 60% Arabic script	2,244
no arb_Arab caption	361
degenerate repetition	121

The repetition filter matters: NLLB collapses into token loops on hard inputs, producing captions like ضوْ ضوْ ضوْ ضوْ .... Rows are dropped when unique-token ratio falls below 0.5 or any token repeats three times consecutively.

No score threshold was applied. The distribution is concentrated between 4.5 and 5.5, so it does not separate good captions from bad ones.

Caveat on quality

These captions are doubly machine-generated: BLIP wrote English captions for LAION-COCO, then NLLB-200 translated them to Arabic. They are usable training signal but not gold references, which caps how much reference-based metrics like chrF or BLEU can tell you.

Reproduce

arabic_vlm/data_prep.py and arabic_vlm/download.py in the model repo.

License

cc-by-nc-4.0, inherited from the source dataset. Non-commercial use only.