eilamc14/wikilarge-clean

WikiLarge Cleaned

Summary
This dataset is a cleaned and deduplicated subset of the classic WikiLarge-style sentence pairs (English Wikipedia → Simple English Wikipedia).
Starting from the original alignment files (wiki.full.aner.ori.train/valid/test.{src,dst}), we constructed a Hugging Face datasets corpus, applied a set of cheap filters, and removed near-duplicates.

Provenance & License: This is a derivative of Wikipedia / Simple English Wikipedia content under CC BY-SA.
The derivative is distributed under CC BY-SA 4.0 as required by the ShareAlike clause.
Please ensure proper attribution (TASL: Title, Author(s), Source, License) and preserve the same license for downstream derivatives.

• Original alignment resource (WikiLarge): DRESS GitHub
• This dataset (code & scripts) can be found in the repository: Simplify This Github

---

Data Schema

• Fields

  • source (str) – original (complex) sentence from English Wikipedia
  • target (str) – simplified sentence from Simple English Wikipedia

• Splits

  • train, validation, test (derived from the original train/valid/test files)

---

Construction & Cleaning

Starting point
We downloaded the original WikiLarge files from the official GitHub and created a standard HF DatasetDict:

• Files used: wiki.full.aner.ori.train.src, wiki.full.aner.ori.train.dst, and analogous valid / test.

Cheap filters (single batched pass) — thresholds used:

• Length constraints
  • Source tokens: MIN_SRC_TOKENS = 4, MAX_SRC_TOKENS = 256
  • Target tokens: MIN_TGT_TOKENS = 2, MAX_TGT_TOKENS = 256
• Compression ratio (simple/complex): keep if 0.40 ≤ CR ≤ 0.95
• Near-identity cutoff (lexical Jaccard): keep if Jaccard(source, target) < 0.98

Implementation notes (as used here):

• Tokenization for checks is whitespace split (str.split()).
• Jaccard is computed on token sets, i.e., lexical overlap.
• We record a per-example first-fail reason among: length → compression → near_identity.

Deduplication strategy.
We deduplicate (source, target) pairs using a stable hash key:

• Normalize each text by lowercasing and collapsing whitespace.
• Compute md5(src_norm) + "#" + md5(tgt_norm); keep the first occurrence and drop repeats.

---

Cleaning Results & Statistics

Sizes

• Train: [296402 → 123862]
• Validation: [992 → 417]
• Test: [359 → 121]

Removal breakdown (exclusive, first-fail)

• Train

  • length fails: [15764] (5.32%)

  • compression fails: [156027] (57.84%)

  • near_identity fails: [633] (0.93%)

  • dedup: [116] (0.04%)

  • before → after

    • Mean CR: [0.88 → 0.70]
    • Mean source length: [25.17 → 26.49]
    • Mean target length: [18.51 → 18.29]

• Validation

  • length fails: [57] (5.75%)
  • compression fails: [572] (57.66%)
  • near_identity fails: [16] (1.61%)
  • dedup: [0] (0.00%)

• Test

  • length fails: [0] (0.00%)
  • compression fails: [237] (66.02%)
  • near_identity fails: [55] (15.32%)
  • dedup: [0] (0.00%)

---

Intended Use

• Primary task: Text simplification (English).
• Recommended splits: Use validation for model selection and test for final reporting.
• Caveats: Original sentence alignments were automatically mined; occasional noise or content drift may exist even after this cleaning.

---

Usage

from datasets import load_dataset

ds = load_dataset("eilamc14/wikilarge-clean")
print(ds)
print(ds["train"][0])  # {"source": "...", "target": "..."}

---

Attribution & License

• Provenance: Derived from English Wikipedia and Simple English Wikipedia content.
• License: CC BY-SA 4.0 (ShareAlike applies). You must provide attribution (TASL: Title, Author(s), Source, License) and distribute derivatives under the same license.

---

Reproducibility

Filters and dedup are implemented as described above (length, CR, Jaccard, and hash-based pair dedup).
Implementation reference: dataset_cleaning.ipynb

---

Citation

Please cite both the widely used WikiLarge reference and this cleaned subset:

Original alignment (WikiLarge): Zhang, X., & Lapata, M. (2017). Sentence Simplification with Deep Reinforcement Learning. In EMNLP 2017 (pp. 584–594). Association for Computational Linguistics. doi:10.18653/v1/D17-1062

This dataset (cleaned subset):
Cohen, E., Bul, I., Inbar, D., & Loewenbach, O. (2026). Simplify-This: A Comparative Analysis of Prompt-Based and Fine-Tuned LLMs. arXiv. arxiv.org/abs/2601.05794

WikiLarge BibTeX:

@inproceedings{zhang-lapata-2017-sentence,
  title     = {Sentence Simplification with Deep Reinforcement Learning},
  author    = {Zhang, Xingxing and Lapata, Mirella},
  booktitle = {Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing},
  pages     = {584--594},
  year      = {2017},
  address   = {Copenhagen, Denmark},
  publisher = {Association for Computational Linguistics},
  doi       = {10.18653/v1/D17-1062},
  url       = {https://aclanthology.org/D17-1062/}
}

Wikilarge cleaned BibTeX:

@misc{simplifythis2025,
  author       = {Cohen, Eilam and others},
  title        = {Simplify-This: A Comparative Analysis of Prompt-Based and Fine-Tuned LLMs},
  year         = {2025},
  howpublished = {\url{https://github.com/eilamc14/Simplify-This}},
  note         = {GitHub repository},
  urldate      = {2025-09-30} 
}