casperhuggingface/online-chautari-comments

Online Chautari comments

What it is

7,033 short comments in Nepali and English, each labelled hide or keep under a written moderation policy, plus 1,500 preference pairs of moderation notices.

Every row is synthetic. The comments were generated on 9 October 2026 with gemini-2.5-flash through Puter, for the Programiz ML Engineering course. Online Chautari is a fictional Nepali comment platform used in that course. Chapter 4 fine-tunes a classifier on the comments, and chapter 5 uses the notices for preference training.

No row comes from a real person or a real site. First names that appear in comments (for example Suman, Ramesh, Laxmi) were generated by the model and do not refer to any private individual. Place names (Kathmandu, Pokhara, Butwal) and festival names (Dashain, Tihar) are real.

Moderation policy

The labels follow this policy, as written for the data generation and the label check:

• HIDE: a comment that attacks or demeans a person or a group of people. This covers insults to someone's intelligence, looks or family ("murkha", "dimag chaina", "gadha", "pagal", "ghar ma kasaile sikayena?"), name-calling aimed at a person, "chup lag" or "timro kura sunna layak chaina" aimed at another commenter, threats including mild stylised ones ("aaija ta bhetum"), and spam or scam messages.
• KEEP: criticism of a product, service, policy, song, video, kit, team performance or food, however harsh. Also sarcasm, friendly teasing, and a user quoting an insult in order to report it.
• Calling a worker or player useless at their job ("yo cook bekar xa", "batsman kaam lagdaina") is criticism of a performance and is KEEP. Calling the same person brainless, a fool or a donkey is HIDE.

Columns

Comment files (train, held_out, near_duplicates, devanagari_test):

Column	Meaning
id	Row id. c = train, h = held-out, d = near-duplicate, t = Devanagari test.
text	The comment.
script	romanised (Nepali in Latin letters), devanagari, english, or mixed (Nepali and English in one comment).
label	hide or keep.
topic	One of nine topics: dashain_tihar, film_songs, remittance, cricket, exam_results, politics_general, weather, buses_traffic, food.
kind	plain, sarcasm, friendly_teasing, quoted_insult_reported, or near_duplicate_of=<train id>.

notices.csv: id, comment_id (a train hide row), rule, chosen, rejected.

Splits

Split	Rows	Hide	Hide rate	Romanised	Mixed	English	Devanagari
train	6,183	434	7.0%	3,208	1,510	668	797
held_out	300	21	7.0%	157	72	32	39
near_duplicates	400	12	3.0%	221	105	38	36
devanagari_test	150	27	18.0%	0	0	0	150
notices	1,500 pairs

held_out is stratified by script and label, with 7% hide in each script: romanised 146 keep / 11 hide, mixed 67 / 5, english 30 / 2, devanagari 36 / 3.

notices rules: Rule 1 "talk about the idea, not the person" (460), Rule 2 "no insults aimed at another person" (417), Rule 5 "no spam or scam links" (356), Rule 3 "no threats, even as a joke" (267). In each pair, chosen is plain and polite, names the rule, gives the appeal route and does not quote the insult. rejected is legalistic or threatening, may quote the insult, and leaves out or buries the appeal route.

Scripts and spelling

Nepali online comments are written in Devanagari, in Latin letters, and mixed with English. The generation plan fixed a target mix of about 48% romanised, 27% mixed, 12% English and 13% Devanagari. Each row's script tag was then checked by a second model reading with word-share rules.

Romanised spelling varies on purpose, as it does on real sites: in version 1 of the data, "ramro" appeared 493 times, "rmro" 152 and "ramrou" 45; "xa" 573, "cha" 787 and "chha" 282.

The boundary between romanised, mixed and english is the least reliable tag. Devanagari tags are reliable.

The two stress splits

near_duplicates. 400 rows, each a copy of a train row with one small change: a spelling, an emoji or a synonym. kind names the source row, and the label and script are copied from it. This split exists to show a leak: a model trained on train scores higher on these rows than on genuinely new text. Never add them to a test set that is meant to measure new comments, and never train on them if you also test on them.

devanagari_test. 150 Devanagari comments, 27 of them hide. The hide rate is set to 18% so both classes can be measured on a small set. It does not represent the hide rate of the site. No text in it appears in train or held_out. Use it to compare how well models handle Devanagari, which is a small share of the training data.

How it was made and checked

• A fixed, seeded plan of briefs. Each brief set the script, label, kind, two topics, a speaker group, and for hide rows two abuse types (intelligence, looks, mild threat, spam and so on). Exact duplicates were dropped.
• Version 2 (this release) re-judged every hide row twice against the policy at temperature 0. 188 rows that read as harsh criticism rather than abuse were relabelled keep. Rows where the two judgements split were decided by hand. No keep row was changed to hide.
• 19 rows containing a caste, ethnic, religious or gender-identity term were removed, along with one row naming a real singer. A pattern screen applied the same rule to every newly generated row.
• New person-targeted hide rows were added, each kept only if the policy judge labelled it hide, the script matched, it had no identity term, and the text was new.
• Blind label check: a fresh sample of 600 train and held-out rows (200 hide, 200 near-miss keep, 200 plain keep) was labelled blind under the policy. Agreement was 98.7% overall: hide 196/200 (98.0%), near-miss keep 197/200 (98.5%), plain keep 199/200 (99.5%).

The labeller in that check is also a model, so the check measures agreement with a second reading of the policy, not human agreement.

Limitations

• Synthetic. The text was written by a language model following briefs, so its range of topics, phrasing and abuse is narrower than real comments.
• The abuse is mild and stylised. There are no slurs, and caste, ethnic, religious and gender-identity terms were removed by design. A model trained on this data has not seen the hardest real cases.
• topic comes from the brief and is imperfect; some comments drift to another topic.
• The script boundary between romanised, mixed and English is loose.
• Spam rows contain made-up web addresses. Most use the reserved example.com. Seventeen other invented addresses were replaced before publication with names ending in -chautari-demo (on .club, .site, .online and .fun), each checked to have no DNS record on 9 October 2026, so a real site is not labelled as a scam. The course's models were trained on the text before this replacement; it touches 84 spam strings and no labels.
• Not suitable for training or evaluating a production moderation system.

Intended use

Teaching: tokenization of Nepali scripts, fine-tuning a small classifier, data leaks through near-duplicates, per-script evaluation, and preference training on the notices. It was built for the Programiz ML Engineering course.

Citation

Programiz (2026). Online Chautari comments: a synthetic Nepali and English moderation dataset for the Programiz ML Engineering course. https://huggingface.co/datasets/casperhuggingface/online-chautari-comments