---
dataset_info:
  features:
  - name: input_ids
    sequence: int32
  splits:
  - name: train
    num_bytes: 35168157552.0
    num_examples: 1073116
  download_size: 18303136476
  dataset_size: 35168157552.0
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
---

# OpenWebTextCorpus tokenized for Gemma

This dataset is a pre-tokenized version of the [Skylion007/openwebtext](https://huggingface.co/datasets/Skylion007/openwebtext) dataset
using the [gemma](https://huggingface.co/google/gemma-2b) tokenizer. As such, this dataset follows the same licensing as the original openwebtext dataset.

This pre-tokenization is done as a performance optimization for using the openwebtext dataset with a Gemma model (gemma-2b, gemma-2b-it, gemma-7b, gemma-7b-it).
This dataset was created using [SAELens](https://github.com/jbloomAus/SAELens), with the following settings:

- context_size: 8192
- shuffled: true
- begin_batch_token: "bos",
- begin_sequence_token: null,
- sequence_separator_token: "bos"
- sae_lens_version: "3.3.0"