iasudakov/owt-gpt2

OpenWebText GPT-2 tokenized cache

This repository contains the prepared OpenWebText cache used for MDLM-MMD training and validation. The Parquet export preserves the original token IDs, attention masks, split membership, and row order. It does not retokenize, filter, or reshuffle the cached data.

Split	Rows	Tokens per row
train	8,730,826	1,024
validation	110,397	1,024

Loading

from datasets import load_dataset

dataset = load_dataset("iasudakov/owt-gpt2")

Use streaming=True to iterate without downloading the entire dataset first. To create a Hugging Face Datasets disk cache, call dataset["train"].save_to_disk(...) and dataset["validation"].save_to_disk(...) separately.

Columns and preprocessing

• input_ids: sequences of GPT-2 token IDs stored as int32.
• attention_mask: sequences stored as float32, retained from the training cache.

The source is OpenWebText, revision b4325f019c648b1641a1784748667e8b74e5e064. The preprocessing code reserves the last 100,000 source documents for validation and uses the preceding documents for training. It tokenizes with gpt2, appends EOS to each document, concatenates token streams within preprocessing batches, and divides them into 1,022-token payloads. Each payload is wrapped with BOS and EOS (GPT-2 token ID 50256), giving 1,024 tokens per row. Remainders within preprocessing batches are dropped.

These row counts refer to packed sequences, not the original document counts. Packing boundaries are specific to this cache; another GPT-2-tokenized OpenWebText dataset is not necessarily an exact replacement.

Export verification and provenance

provenance.json records the source cache fingerprints, exact row counts, contiguous shard ranges, Parquet SHA-256 checksums, and hashes of each shard's first and last rows. Export checks verify the schema, row count, and boundary rows of every shard against the source cache. The source data's content and applicable terms remain those of OpenWebText and its underlying documents.