A pause-injected variant of JackHsieh/statML-arxiv-40M-20M-llama32:
every document token is preceded by a sentinel-wrapped stretch of pause tokens, so each row is exactly
8x longer (4_096 -> 32_768 tokens). Nothing else changes — same papers, same
windows, same train/test split, same schema and column order as the parent.
For each row of the parent, the token sequence [t0, t1, ...] becomes:
<|reserved_special_token_1|> <|reserved_special_token_0|> x5… See the full description on the dataset page:
https://huggingface.co/datasets/JackHsieh/statML-arxiv-40M-20M-llama32-pause7-injected.
