fhai50032/Hindi_corpus-4096-packed-qtk-1.37M

Packed pretraining corpus, 1,369,441 rows x 4096 tokens = 5.609B tokens, tokenized with fhai50032/QTK-81K.

Format

column	type	notes
label	list<int32>	exactly 4096 tokens, no padding
raw_label	string	label decoded back to text (redundant, for inspection)

There is no attention_mask column: the corpus is packed, so every position is a real token and the mask would be all ones on every row.

How it was built

Documents are tokenized and appended to a running buffer with a single <|im_end|> (id 81914) between them. Every time the buffer reaches 4096 tokens a row is emitted and the remainder carries into the next row. So no row is padded, no document is truncated, and a document longer than 4096 tokens simply spans rows.

Source: /home/marimo/work/data/raw/hi.

Important: do not mask id 0

QTK-81K has pad_token_id == eos_token_id == 0. The separator here is <|im_end|> (81914), not <|endoftext|>, so id 0 never appears in this corpus. Masking it is therefore a no-op and is safe, but it is also unnecessary -- there is no padding to mask.

If you adapt this recipe to a corpus that separates documents with <|endoftext|>, masking id 0 would delete every document-boundary target and the model would never learn to stop -- a silent failure that reads as degeneration rather than a bug.