Datasets:
license: cc0-1.0
language:
- hi
- en
tags:
- hinglish
- pretraining
- reddit
size_categories:
- 1B<n<10B
Chad pretrain corpus
The pretraining corpus for Chad, a small from scratch Hinglish chat model. Around 4.13 billion tokens of romanized Hinglish and Indian English, split into 46 shards, one JSON object per line as {"text": ..., "source": ...}.
Where it came from
- Reddit, about 96 percent. Comments from 214 Indian subreddits, pulled from the public Pushshift per subreddit torrent dumps. Only Indian communities were selected so it stays on language and topic.
- Hugging Face romanized Hindi sets (diwank/hinglish-dump), about 2.5 percent.
- A handful of public Hugging Face and Kaggle Hinglish sets (chat pairs, instructions, tweets), about 1.5 percent.
How it was cleaned
Every document went through one filter. First strip the junk: html entities, leftover markup, citation and removed markers, Reddit user refs, quote lines, repeated character runs. Then drop whole documents that are non Latin script (Tamil, Telugu, Devanagari and so on), bot or mod boilerplate, too short, or symbol spam. Finally remove exact duplicates with a Bloom filter, which dropped about 6.5 million identical copies.
Roughly 129 million documents went in and about 4.13 billion clean tokens came out, single language romanized Hinglish only.
No private chat data (WhatsApp, Instagram) is in here. That was kept aside for fine tuning and never published.