Delta351/pg19-icr-data-v1

PG-19 ICR Data v1

Processed full-corpus PG-19 package for controlled causal-language-model architecture experiments.

Source

Official PG-19 benchmark: https://github.com/google-deepmind/pg19

Official public bucket: gs://deepmind-gutenberg

Books:

• train: 28,602
• validation: 50
• test: 100

Tokenizer

• Byte-level BPE
• vocab size: 8,192
• trained on the training split only
• special token: <pad>
• pad id: 0

Packed format

• context length: 1024
• stored row width: 1025
• dtype: uint16
• rows never cross book boundaries
• consecutive rows from the same book overlap by one token
• each within-book next-token target is represented exactly once
• *.book_ids.int32.bin stores the dense source-book ID for every row
• *.valid_targets.uint16.bin stores the number of valid next-token targets

Processed target counts

• train: 3,281,210,962
• validation: 4,900,670
• test: 11,552,848

See manifest.json for exact shard sizes, SHA-256 checksums and preprocessing metadata.