CS230-lol/JigsawToxicity

Jigsaw Unintended Bias — Threaded Toxicity

Comments from the Jigsaw Unintended Bias in Toxicity Classification competition (Civil Comments platform, 2015–2017), with each comment's reply chain reconstructed so a toxicity classifier can see the conversation the comment was written in, not only the comment alone.

Built to fine-tune Kev (an open System One-style decision model) for context-aware toxic chat detection.

Files / configs

Config	File	Rows	Description
threads (default)	threads.jsonl	1,999,516	One record per comment with its ancestor chain as context.
filtered	filtered.csv	1,999,516	all_data.csv reduced to 14 columns (text, thread keys, toxicity labels).
all_data	all_data.csv	1,999,516	Original Kaggle all_data.csv, unmodified (46 columns incl. identity and reaction columns).

from datasets import load_dataset

threads = load_dataset("CS230-lol/JigsawToxicity", "threads", split="train")
train = threads.filter(lambda r: r["split"] == "train")
test = threads.filter(lambda r: r["split"] == "test")

Every config loads as a single train split; the original Kaggle partition is in the split column (train 1,804,875 / test 194,641).

Thread reconstruction

parent_id holds the id of the comment being replied to. Verified on the data:

• 1,134,709 comments are replies (non-empty parent_id); 1,005,497 of those parents exist in the dump.
• Every resolved parent shares the child's article_id.
• The parent is created before the child in all but 2 of 1,005,497 pairs.

For each comment, parent_id is followed upward until a top-level comment (no parent) or a parent missing from the dump. The nearest 5 ancestors are kept, ordered root → direct parent. Sibling replies are not included; context is the linear chain the comment responds to.

threads schema

Field	Type	Description
id	string	Comment id.
split	string	Original Kaggle partition: train or test.
article_id	string	Article the comment thread belongs to.
depth	int	Number of resolvable ancestors (full depth, before the 5-ancestor cap). 0 = top-level or orphan.
parent_missing	bool	True if the chain breaks: some ancestor's parent_id points to a comment not in the dump.
context	list of {id, text}	Up to 5 nearest ancestors, root → direct parent. Empty for top-level comments.
comment	string	The comment to classify.
toxicity	float	Fraction of annotators rating the comment toxic (0–1). Primary target; ≥ 0.5 is the competition's positive class.
severe_toxicity, obscene, sexual_explicit, identity_attack, insult, threat	float	Subtype annotator fractions (0–1).
toxicity_annotator_count	int	Number of annotators for the toxicity labels.

Example (context texts truncated):

{"id": "5167187", "split": "train", "article_id": "328935", "depth": 3, "parent_missing": false,
 "context": [{"id": "5166075", "text": "I wonder how many people have actually been inconvenienced since this initiative took effect..."},
             {"id": "5166798", "text": "Does it make a difference? Is that the standard for right and wrong?..."},
             {"id": "5166939", "text": "So from what I gather from your comment, you do not view transgender people as normal..."}],
 "comment": "That's right. They are not normal. And I am starting from the premise that they are ABNORMAL. Proceed wth the typical racist, bigot, sexist rubbish.\nThanks!",
 "toxicity": 0.803, "severe_toxicity": 0.066, "obscene": 0.049, "sexual_explicit": 0.082,
 "identity_attack": 0.131, "insult": 0.754, "threat": 0.016, "toxicity_annotator_count": 61}

The reply alone reads as an insult; the context shows it is aimed at transgender people, which is the kind of signal context is meant to supply.

filtered columns

id, comment_text, split, created_date, parent_id, article_id, toxicity, severe_toxicity, obscene, sexual_explicit, identity_attack, insult, threat, toxicity_annotator_count. parent_id is normalised from float ("123.0") to an integer string.

Statistics (threads)

	Count
Records	1,999,516
With non-empty context	1,005,497 (50.3%)
parent_missing = true	154,182
toxicity ≥ 0.5	159,782 (8.0%)
toxicity ≥ 0.5 with context	75,626
Articles	116,943

Depth of replies with a resolvable parent: 1 → 441,276; 2 → 239,914; 3 → 129,782; 4 → 72,367; 5 → 42,326; 6–9 → 60,823; ≥ 10 → 19,009.

Preprocessing

Reproducible with two standard-library Python scripts (no dependencies):

python3 filter.py        # all_data.csv -> filtered.csv
python3 build_threads.py # filtered.csv -> threads.jsonl

Limitations and caveats

• Split leakage across threads. The Kaggle train/test partition is per comment, not per thread, so a test comment's context can contain train comments (and vice versa). For leakage-free evaluation, re-split by article_id.
• Incomplete chains. About 11% of replies point to a parent absent from the dump; parent_missing flags these. Their context is partial or empty, which does not mean they are top-level comments.
• Context is truncated to 5 ancestors; long chains lose their earliest turns. Text length is not truncated.
• Identity columns dropped from filtered/threads (mostly empty). Use all_data to compute the competition's unintended-bias metrics (subgroup AUC, BPSN, BNSP).
• Class imbalance. Only 8% of comments are toxic at the 0.5 threshold.
• Offensive content. The data contains real toxic, obscene, threatening and identity-attacking text.
• Labels are crowd-annotator fractions and contain annotator disagreement and bias.

License and citation

Civil Comments text and Jigsaw annotations are released under CC0 1.0. The thread reconstruction adds no new text.

@misc{jigsaw-unintended-bias-in-toxicity-classification,
  author = {cjadams and Daniel Borkan and inversion and Jeffrey Sorensen and Lucas Dixon and Lucy Vasserman and nithum},
  title = {Jigsaw Unintended Bias in Toxicity Classification},
  year = {2019},
  howpublished = {\url{https://kaggle.com/competitions/jigsaw-unintended-bias-in-toxicity-classification}},
  publisher = {Kaggle}
}

@article{borkan2019nuanced,
  title = {Nuanced Metrics for Measuring Unintended Bias with Real Data for Text Classification},
  author = {Borkan, Daniel and Dixon, Lucas and Sorensen, Jeffrey and Thain, Nithum and Vasserman, Lucy},
  journal = {arXiv preprint arXiv:1903.04561},
  year = {2019}
}