ahmedheakl/arocrbench_khattparagraph

Please see paper & code for more information:

• https://github.com/mbzuai-oryx/KITAB-Bench
• https://arxiv.org/abs/2502.14949

Ground truth correction (2026-08-10)

The answer column previously contained only the first line of each handwritten paragraph rather than the full text (mean 68 characters against 336). Models that correctly transcribed the whole image were penalised for the text the annotation was missing, which produced CER and WER values above 4.0 on this subset. Reported in issue #5.

The images were never affected. Full paragraph transcriptions were recovered by matching all 200 images pixel-for-pixel (MD5 over decoded greyscale) against a-alnaggar/khatt-paragraphs, the processed KHATT paragraph set. All 200 matched exactly, with no hash collisions among that dataset's 3,996 images.

The recovered text was then normalised: the transcribers' # line-break marker was removed, tatweel and zero-width/bidi characters were stripped, line breaks were folded into spaces, spacing around punctuation was regularised, and repeated periods were collapsed to a single ellipsis. Percent signs and digits were kept, being genuine content (e.g. 70 %, 0.5%), as were the writers' own spelling variants, which are part of the KHATT transcription convention.

Splits

Split	Contents
train	Corrected full-paragraph annotations. Use this.
old	The previous truncated annotations, kept so earlier results stay reproducible.

load_dataset("ahmedheakl/arocrbench_khattparagraph", split="train")  # corrected
load_dataset("ahmedheakl/arocrbench_khattparagraph", split="old")    # as published before

Both splits hold the same 200 images in the same order; only answer differs. Scores computed against old are not comparable with scores computed against train.