compling/cs489-f26-a2

A2 English data

This dataset contains cleaned English Wikipedia training and development text for CS 489 Assignment 2 at the University of Waterloo (Fall 2026). The public release is compling/cs489-f26-a2, part of the CS 489 F26 collection. Only train and dev are released; no held-out evaluation inputs are included.

Split	Documents	Cleaned document characters	TXT characters	TXT UTF-8 bytes
Train	854	10,007,545	10,009,252	10,057,865
Dev	76	1,014,702	1,014,853	1,021,824

Load with Hugging Face Datasets

Install datasets and load the default configuration:

from datasets import load_dataset

data = load_dataset("compling/cs489-f26-a2")
assert len(data["train"]) == 854
assert len(data["dev"]) == 76
article = data["train"][0]
text = article["text"]

The split names are train and dev (not validation). Each row is one complete article, in the original JSONL order, with all 18 text and provenance fields preserved. Page IDs, revision IDs, and revision timestamps are strings; character and byte counts are integers. The explicit Parquet split configuration prevents the original TXT and JSONL copies from being loaded as additional examples or split into individual lines. Parquet preserves the JSONL field types exactly, including the original ISO 8601 revision timestamp strings.

For the assignment's plain-text input files, download the original TXT files without reconstructing document boundaries:

from huggingface_hub import hf_hub_download

for split in ("train", "dev"):
    hf_hub_download(
        repo_id="compling/cs489-f26-a2",
        filename=f"{split}.txt",
        repo_type="dataset",
        local_dir="data",
    )

All TXT and JSONL files are byte-identical to the prepared release and match manifest.json. For reproducible experiments, supply the same dataset commit hash as revision= to both APIs.

Files and boundaries

• train.parquet and dev.parquet: Hugging Face dataset tables, preserving every field and row from the corresponding JSONL.
• train.txt and dev.txt: UTF-8 text without a BOM, complete articles separated by two newline characters, followed by one final newline.
• train.jsonl and dev.jsonl: one complete article per JSON line, in exactly the corresponding TXT order.
• manifest.json: pinned source and cleaner provenance, exact counts, and SHA-256 checksums.
• LICENSE.txt: license, attribution, and modification notice.

Every JSONL record includes id, title, url, and text. Additional fields identify the source snapshot, exact Wikipedia revision and timestamp, contributor-history URL, license/attribution, raw and cleaned text hashes/counts, and a whitespace-normalized deduplication hash. Use JSONL when explicit document boundaries or attribution are needed: blank lines inside articles make TXT blank-line splitting ambiguous. Parse JSONL one physical line at a time and decode each line as JSON.

The character quotas count Unicode code points in cleaned document text, not UTF-8 bytes or tokenizer tokens. The TXT character and byte totals include inter-document separators and the final newline. Whole documents are retained even when a split slightly exceeds its quota. Train and dev have disjoint page IDs, URLs, cleaned-text hashes, and whitespace-normalized cleaned-text hashes. No vocabulary is supplied or inferred from held-out data; a fixed UTF-8 byte tokenizer requires no corpus-derived vocabulary.

Provenance and caveats

Source: English Wikipedia, enwiki-20260901 raw XML dump. Only a bounded compressed prefix was retrieved, not the complete dump. The original input was raw wikitext, not an upstream plaintext extraction. Each full article was cleaned by the unmodified CS 489 A1 standard clean_text(raw, wiki='en') implementation. Its SHA-256 is 79bfc0ed7cfb32ba167d7d5b42bdfb1a6e1871a610cb1894d1e4bbccef020626. See the manifest for snapshot, source-checksum, and preparation-code details.

This is a shuffled, filtered prefix sample, not a uniform Wikipedia or general-English sample. Early page order favors older articles, and the 500 to 30,000 cleaned-character length bounds exclude short stubs and long articles. Alphabetic/Latin-letter filters, Wikipedia editorial biases, markup removal, and residual extraction artifacts affect the text. Related articles and repeated passages can remain despite whole-document deduplication. Source Unicode punctuation is preserved; character counts and byte counts therefore differ.

License and attribution

Copyright belongs to the English Wikipedia contributors. The cleaned adaptation is supplied under Creative Commons Attribution-ShareAlike 4.0. Article, exact-revision, and page-history links in JSONL provide per-document attribution and contributor histories. See Wikipedia:Copyrights for individual notices and additional licensing conditions; GFDL eligibility is subject to exceptions. Modifications consist of A1 markup removal/normalization, whole-document filtering and deduplication, reordering, and TXT separators. Keep the corresponding JSONL, LICENSE.txt, and manifest with redistributed text, retain attribution, and comply with the applicable share-alike terms.