robloxianer/ShitpostGPT-corpus

ShitpostGPT corpus (recipe)

The training data behind ShitpostGPT-10M and ShitpostGPT-20M: about 915 MB of text (about 241M tokens) mixing clean web text with Twitch chat, Reddit shitposting comments and imageboard posts.

This repo contains no raw text. Almost all of the corpus is user-generated content from other people's datasets, and the licenses on those range from unclear to none, so I'm not re-hosting it. Instead this repo has the exact recipe: the script that streams the public source datasets, cleans them and mixes them into the same corpus, plus the stats and the tokenizer. Anyone can rebuild it in about 10 to 20 minutes.

What is in this repo

file	what it is
README.md	This card
build_corpus.py	The corpus builder (streams the sources, cleans, dedupes, bundles, interleaves, writes shards)
stats.json	The real per-source sizes of the corpus used for ShitpostGPT-20M
tokenizer.model, tokenizer.vocab	The 16,384-token SentencePiece BPE tokenizer trained on the corpus

Composition (the ShitpostGPT-20M corpus)

source	what it is	size	share
HuggingFaceFW/fineweb-edu (config sample-10BT)	Clean English web text, so the model learns grammar	450 MB	49%
simonko912/chan-shitpost-2.5	Imageboard posts	248 MB	27%
innovation-hacking2/shitposting_dataset0 to 5	Reddit shitposting comments	119 MB	13%
lparkourer10/twitch_chat	Twitch chat messages	82 MB	9%
DarwinAnim8or/greentext	Greentext stories	0.3 MB	under 1%
A small personal scrape (not included)	Twitch chat and Reddit comments, collected from public sources	17 MB	2%
Total		915 MB

Without the personal scrape the rebuilt corpus is about 899 MB. The ShitpostGPT-10M corpus was a slightly smaller mix of the same sources (about 620 MB, 8,192-token tokenizer).

Tokenized with the included tokenizer, the 20M corpus is 240.5M training tokens and 0.93M validation tokens (every 200th chunk goes to validation).

How the corpus is built

build_corpus.py does the following for every source:

1. Reads only the text column. No usernames, timestamps or other metadata are read or kept.
2. Cleans each text: HTML entities unescaped, <|endoftext|> markers removed, URLs removed, imageboard reply references (>>12345) removed, whitespace collapsed, empty lines trimmed.
3. Filters by length (in characters): fineweb 200 to 6000, Twitch 4 to 500, imageboard 5 to 4000, Reddit 8 to 3000, greentext 20 to 4000.
4. Removes exact duplicates across all sources (hash of the cleaned text).
5. Bundles short lines into threads. For the short sources (Twitch, imageboard, Reddit, personal scrape), 3 to 15 consecutive lines (300 to 1500 characters) are joined with newlines into one document. This teaches a model to keep writing multi-line chat instead of stopping after one short line. Fineweb and greentext are not bundled.
6. Interleaves all sources by weighted random choice, so every training window is a mix. The random seed is 1337.
7. Writes shards of about 100 MB as JSON lines: {"t": "<text>", "s": "<source>"}.

The model's training code tokenizes each document as [BOS] + tokens + [EOS] and packs them into one stream.

Rebuild it

pip install datasets huggingface_hub
python build_corpus.py                  # full corpus into ./corpus/
python build_corpus.py --scale 0.01     # 1% test run
python build_corpus.py --no-bundle      # keep short lines as separate documents

Your own extra text file can be added with --local path/to/lines.txt (one document per line); if the file is missing, that source is skipped. The upstream datasets can change or disappear, so a rebuild may differ slightly from the corpus used for the published models.

Limitations and content warning

• No toxicity filtering. The imageboard and Reddit parts contain slurs, harassment, sexual content, conspiracy talk and other offensive text. That is intentional for a shitposting model, but it means models trained on this data will reproduce it.
• The text is user-generated and may contain personal information despite the URL removal. I did not run any PII scrubbing.
• English only, and heavily skewed toward gaming and internet culture.
• Not suitable for training assistants, moderation tools or anything user-facing without heavy filtering.

Licenses

This repo's script and card are released as-is with no warranty. The data itself is not included: each source dataset has its own license and terms, which you have to check on its page before downloading or using it. For example, FineWeb-Edu is published under the Open Data Commons Attribution License. The Reddit, Twitch and imageboard datasets may carry additional terms from the platforms they come from.

Related

• Models: ShitpostGPT-20M (and ShitpostGPT-10M)
• Credit to Codeically's "ChatGPT-2 on a potato" video for the original idea.