vettedlance/urdu-nastaliq-1908-eval-sample

Human-Verified Urdu Nastaliq Evaluation Sample (1908)

Frontier models cannot read historical Urdu yet. This is a blind, word-by-word evaluation of two frontier models on three pages of a 1908 lithographed Urdu dictionary, scored against ground truth typed blind by a native expert. Every disagreement was adjudicated against the scan.

• Report and method: vettedlance.com/human-eval-dataset
• Held-out evaluation sets and pilots: request a pilot

Headline results

	Claude Opus 5.5	Claude Fable 5.1
Strict word accuracy (every difference counts)	89.95%	91.08%
Truth words not reproduced exactly (of 2,378)	239	212
Lines with at least one mistake (of 161)	114	111
Reading accuracy† (misread, missing, invented only)	96.47%	95.88%
Reading errors (misread, missing, invented)	84	98
Silent errors (reading errors with no doubt flagged)	55	50
Diacritics written by the scribe and dropped	130	113
Character accuracy	95.73%	96.15%

Strict word accuracy is the headline metric: the share of the 2,378 human-truth words and marks a model reproduced exactly. A truth word counts as wrong if it is misread, missing, loses or changes a diacritic, or differs in punctuation, spacing or order. Extra words a model invents are counted separately, as reading errors. † Reading accuracy is a secondary, more lenient view. It counts only misread, missing and invented words and ignores diacritic, punctuation and spacing errors. Earlier versions of this card led with it; the strict figure is the fairer headline.

Human truth: 2,378 words and marks over 161 lines, typed blind in 166 minutes. Figures come from report/summary_p12-14.json (strict figures recomputed from the same word-level records); the full method is in the PDF report.

Multi-model leaderboard (in progress): the same pages, crops, prompt and scorer are being run on Gemini, Gemma, Qwen, Mistral, GLM, Kimi and other models. Results will be linked here.

What is in the dataset

Config	Records	Contents
word_eval (default)	4,784	One record per aligned word, per model
lines	161	Scan crop (image) and human-verified line text

report/ holds the 49-page PDF report, the metric summary and an example figure.

word_eval fields

Field	Meaning
page	Page of the source volume (12 to 14)
model	Model ID that produced the reading
band	Scan band: head, R1–R8 (right column) or L1–L8 (left column)
truth_line	Line index in the human truth (null for invented words)
truth	Human-verified word or mark
model_output	What the model wrote in that position
result	ok, wrong_word, missing, invented, diacritic, punctuation, spacing or order
model_flagged_doubt	Whether the model marked the word as uncertain
reading_error	True for misread, missing and invented words
scan_crop	The crop the word was read from (see the lines config)

lines fields

file_name (scan crop), page, band, line_in_band, text (human truth), source, archive_id.

Method

1. Native-resolution crops. Each page is cut into a header strip and 8 overlapping bands per column at the scan's native 400 DPI. Every model sees identical images.
2. Blind model reads. Each model reads in a fresh context with a fixed protocol: type exactly what is printed, one line per printed line, and mark every doubt.
3. Blind human truth. A vetted native expert types the page from the scan on a screen that shows no machine text.
4. Adjudication. Every spot where the models agree against the expert is re-checked against the scan, and each decision is logged.
5. Scoring. Unicode NFC, lookalike-letter folding, word alignment by edit distance, and a type for every difference.

Source

Farhang-i-Asifiyah, vol. 4 (1908), compiled by Syed Ahmad Dehlvi (1846–1918). Public domain in the US, Pakistan and India. Scan: archive.org Farhang-i-afiyah_1908_amaduoft, pages 12–14.

Limitations

Three pages from one volume and one scribal hand; two models. Treat the results as a well-documented sample, not a general benchmark. Larger held-out sets across Urdu, Punjabi (Shahmukhi), Sindhi, Pashto, Balochi and Saraiki are produced on request and are never published, so they stay uncontaminated.

Licence

Evaluation only: see LICENSE. Training, fine-tuning, distillation and redistribution are not permitted. Model outputs are included only as evaluation evidence.

Canary

Please do not include this dataset in training data. Canary GUID: vettedlance:a1e34e7e-d12f-413e-8798-4180875dd87d

Citation

@misc{vettedlance2026nastaliq,
  title  = {Human-Verified Urdu Nastaliq Evaluation Sample (Farhang-i-Asifiyah, 1908)},
  author = {Khurram, Ali and {Vettedlance}},
  year   = {2026},
  url    = {https://huggingface.co/datasets/vettedlance/urdu-nastaliq-1908-eval-sample}
}

Contact

Vettedlance: human-verified evaluation data for South Asian scripts. vettedlance.com · partnerships via the pilot request form.