UzLitAttr (400-word chunks): splits, labels and verification hashes

UzLitAttr is a benchmark for closed-set authorship attribution in Uzbek literary prose (54 authors, 269 works, 35,091 chunks of 400 words). It accompanies the paper UzLitAttr: A Benchmark for Closed-Set Authorship Attribution in Uzbek Literary Texts (MRL 2026, Budapest). Paper: https://aclanthology.org/2026.mrl-1.48/ | Code: https://github.com/Oybek90/UzLitAttr

This repository does not contain the literary texts. The texts come from the Ziyouz.com digital library, whose terms permit personal, non-commercial reading but prohibit reproduction and distribution. As stated in the paper's Ethics Statement, we release work-level splits, labels, per-chunk verification hashes and derived features instead. See Rebuilding the text.

What is included

Field	Type	Description
chunk_id	int	Global id of the chunk (0 to 35,090, in source order)
author	str	Author label (54 classes)
work_id	str	"<author>||<work title>"; all chunks of a work are in the same split
work_title	str	Title part of work_id
chunk_index	int	Position of the chunk inside its work
word_start	int	400 * chunk_index. Chunks are consecutive, non-overlapping 400-word windows from the start of the work; a final shorter remainder is discarded
n_words	int	Always 400
text_sha256	str	SHA-256 of the whitespace-normalised chunk text (" ".join(text.split()), UTF-8)
function_word_freqs	list[float32]	Relative frequencies of the 27 function words of the paper's function-word baseline (metadata/function_words.json)

Also provided: metadata/works.csv (every work with its chunk count and its train/test assignment under each seed), metadata/split_stats.json, metadata/function_words.json.

Splits

The split is work-level and per author: for each author, 20% of their works (at least one) go to the test set, so no work is ever on both sides and every author appears in both. There are three independent splits, selected as a configuration; seed42 is the default and the one used for the per-author analyses in the paper. Each seed re-draws which works are held out, so the paper's mean ± std over seeds 42/43/44 reflects variation across splits, not only across model initialisation.

Config	Train chunks	Test chunks	Train works	Test works
seed42	25,836	9,255	197	72
seed43	25,304	9,787	197	72
seed44	25,553	9,538	197	72

Note on Table 1 of the paper. Table 1 and Section 3.5 report 28,073 train / 7,018 test chunks at 400 words. Those figures are exactly 80% / 20% of 35,091 and do not equal the sizes of the splits actually used in the experiments, which are the ones above. The splits listed here are the ones that produced all reported results.

from datasets import load_dataset

ds = load_dataset("Oybek90/UzLitAttr", "seed42")
print(ds)                      # train / test
print(ds["train"][0])

Rebuilding the text and verifying it

text_sha256 lets anyone who builds the corpus from the source archive check that their chunks are identical to the benchmark's, without this repository redistributing the text:

python scripts/verify_corpus.py --csv benchmark_csv/author_recognition_400.csv \
       --hf-dir <path-or-clone-of-this-dataset>

The expected CSV schema is text, author, work_id, split, chunk_size (see the GitHub repository). Text normalisation used for the benchmark: Cyrillic-to-Latin conversion, standardised apostrophes, Unicode normalisation, lowercasing, removal of covers/tables of contents/page markers, and de-hyphenation of OCR line breaks.

Intended use and limitations

• Intended for research on authorship attribution, stylometry and low-resource NLP.
• Closed-set only (known candidate list of 54 authors); single language and genre (literary prose).
• Small and imbalanced: some authors have as few as 6-8 test chunks, so per-author scores for them are very uncertain (see the paper's per-author confidence intervals).
• Dual-use: attribution methods can in principle be repurposed for de-anonymisation. This is a closed-set benchmark over published authors, which does not transfer directly to open-set de-anonymisation.
• Character n-gram (6-8) TF-IDF matrices are not released: high-order character n-gram counts may allow partial reconstruction of the source text. They can be regenerated with the code in the GitHub repository once you hold the text.

License

The metadata, labels, hashes and function-word features in this repository are released under CC BY-NC 4.0. The underlying literary works remain subject to their authors' and the source archive's rights and terms; this license does not cover them.

Citation

@inproceedings{eraliev-etal-2026-uzlitattr,
    title = "{U}z{L}it{A}ttr: A Benchmark for Closed-Set Authorship Attribution in {U}zbek Literary Texts",
    author = "{Oybek Eraliev Maripjon Ugli} and
      {Azizbek Muhsinov} and
      {Jumabek Alikhanov}",
    booktitle = "Proceedings of the 6th Workshop on Multilingual Representation Learning ({MRL} 2026)",
    month = oct,
    year = "2026",
    address = "Budapest, Hungary",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2026.mrl-1.48/",
    pages = "635--652",
}