This dataset contains public forum posts by real people. Usernames are pseudonymized and contact details are scrubbed, but the text is otherwise as written. By requesting access you agree to use it for research only, to not attempt to re-identify any author, and to honor removal requests.

Nairaland Conversations

1,901,705 thread pages and 34,187,603 posts from Nairaland, the largest Nigerian discussion forum, with the thread structure kept intact: who posted, when, and who they were quoting. About 6.9 billion characters of Nigerian English, Nigerian Pidgin and code-switched conversation, the kind of text that is badly underrepresented in public training data.


thread pages	1,901,705
posts	34,187,603
characters (post text)	6,914,511,323
distinct (pseudonymous) authors	659,239
parquet shards	12 (6.4 GB, zstd)
built	2026-10-15

Loading

from datasets import load_dataset

ds = load_dataset("caleb-mnb/nairaland-conversations", split="train", streaming=True)
for page in ds.take(1):
    print(page["title"], page["board"])
    for post in page["posts"]:
        print(post["author_id"], post["posted_at"], post["text"][:80])

Structure

One row per thread page (a long thread spans several pages).

field	type	notes
`id`	string	hash of the canonical url
`url`	string	source page, kept for provenance and removal requests
`thread_id`	int	nairaland topic id, null for the few pages reached only via a post permalink
`page`	int	0-based page of the thread
`board`	string	forum section (Politics, Romance, ...), parsed from the page title
`title`	string	thread title
`lang`	string	dominant language of the page, by characters
`n_posts` / `n_chars`	int	post count, total post text length
`crawled_at`	string	date the page was fetched
`text`	string	the whole page flattened as `author_id:` then the post, one block per post
`posts`	list	per post: `index`, `author_id`, `posted_at`, `edited_at`, `lang`, `text`, `quotes`

quotes holds what a post replied to: the quoted author's pseudonym, the nairaland post id that was quoted, and the quoted text. Quoted text is removed from the post's own text, so a post's words stand alone. Timestamps are local time as nairaland shows it (WAT, UTC+1); where the site omitted the year or date, it was filled in from the crawl date.

Languages

Tagged per post with GlotLID, on the first 400 characters. Posts under 20 characters are not tagged. GlotLID often confuses Nigerian Pidgin with other English-based creoles (Jamaican, Cameroonian, Krio), so those labels are folded into pcm_Latn, and predictions outside the languages you'd expect on a Nigerian forum, or below 0.3 confidence, are marked und. Short, code-switched text is hard for any language id model, so treat these as estimates, especially English vs Pidgin.

language	posts	share of text
English (`eng_Latn`)	21,858,755	94.35%
Nigerian Pidgin (`pcm_Latn`)	4,160,039	5.48%
Yoruba (`yor_Latn`)	27,786	0.06%
Igbo (`ibo_Latn`)	35,430	0.06%
Hausa (`hau_Latn`)	8,105	0.01%
Tiv (`tiv_Latn`)	14,631	0.01%
French (`fra_Latn`)	4,454	0.01%
Arabic (`arb_Arab`)	332	0.01%
Efik (`efi_Latn`)	2,854	0.00%
German (`deu_Latn`)	2,189	0.00%
Urhobo (`urh_Latn`)	3,036	0.00%

Boards (top 25 by pages)

board	pages	posts
Politics	358,613	7,435,138
Romance	163,448	3,619,213
Jobs/Vacancies	89,621	1,153,453
Education	83,019	1,380,590
Crime	80,934	1,679,986
Business	76,882	1,113,971
Family	68,657	1,535,579
Christianity Etc	65,480	1,222,104
Celebrities	61,375	1,193,354
Health	59,502	1,021,213
Autos	52,667	799,218
Sports	49,369	860,774
Nairaland General	46,262	502,218
Travel	45,761	1,027,805
Technology Market	43,481	678,107
Foreign Affairs	43,458	892,965
Phones	41,243	683,877
Properties	35,591	492,053
Career	35,298	554,192
Jokes Etc	35,075	508,996
Culture	30,436	596,426
Food	28,568	537,637
Literature	23,609	444,900
TV/Movies	19,935	394,796
Investment	19,556	252,917

Posts per year

year	posts
2005	40,604
2006	139,078
2007	198,232
2008	332,907
2009	426,001
2010	406,337
2011	480,024
2012	881,833
2013	1,385,442
2014	1,118,923
2015	2,535,414
2016	2,344,741
2017	2,305,115
2018	2,106,199
2019	2,278,379
2020	2,982,911
2021	2,540,321
2022	2,760,426
2023	2,506,673
2024	2,183,770
2025	2,183,795
2026	1,444,483

Collection

Crawled with WebHarvester, an open source, resumable, distributed crawler built on Crawlee. The crawl was seeded at 15 discussion boards, stayed on the nairaland domain, and followed links to a depth of 4. It skipped login, posting, search, marketplace and advertising pages, and user profiles. Requests were throttled per domain and honored HTTP 429 backoff.

Cleaning and privacy

Only thread pages are included. Profile pages, login pages and per-user post listings were removed (2,986 pages skipped as empty or unusable).
1,615,150 duplicate pages were removed (mostly post permalinks that resolve to a thread page already in the set).
Usernames are replaced with stable pseudonyms (u_ plus 12 hex characters, a keyed hash). The same user has the same id across the dataset, so conversations and reply chains stay intact, but the key is not published. Every @handle in post text, and bare references to usernames on the same page, are replaced with the same keyed pseudonyms (2,971,047 replacements), so a user mentioned on another page still maps to their one id.
Email addresses (417,241), phone numbers (1,271,990), BlackBerry Messenger PINs (30,061) and 10 to 11 digit numbers such as bank account numbers (91,478) were replaced with [EMAIL], [PHONE], [PIN] and [NUMBER].
Links to social media profiles (Facebook, Twitter/X, Instagram, TikTok, LinkedIn, Telegram, WhatsApp) were replaced with [SOCIAL] (125,873).
Leftover forum markup ([quote], [img] and similar tags, 1,835,549 of them) was stripped, keeping the text between the tags.

This is not perfect anonymization. People mention real names, places and life details in their own posts, and a determined reader could link a pseudonym to a public profile by searching for a post's text. If you find content about yourself that you want removed, open a discussion on this page with the url and it will be dropped from the next version.

Limitations and bias

It reflects who posts on Nairaland: mostly urban, internet-connected, and more male than the general population. Political and religious threads are heated, and the text includes insults, slurs and misinformation as written.
Depth-limited crawl, so recent threads and the first pages of boards are overrepresented relative to the forum's full archive.
Images, embedded media and formatting are not included. Posts that were only an image appear with empty text.
The quote graph points at nairaland post ids, but posts in this dataset don't carry their own id, so a quote can be matched to its source by text and author, not by id.

License

The compilation (the selection, cleaning, pseudonymization and structure) is released under CC BY-NC 4.0. The posts themselves remain the property of their authors, and this license grants no rights to them beyond what the access terms above allow: non-commercial research, with no attempt to re-identify anyone. Nairaland's own terms of service also apply to the original content.

Citation

@misc{olusola2026nairaland,
  title  = {Nairaland Conversations: a Nigerian forum corpus with thread structure},
  author = {Olusola, Caleb},
  year   = {2026},
  url    = {https://huggingface.co/datasets/caleb-mnb/nairaland-conversations}
}

Downloads last month: 42

Total file size:

6.42 GB