|
--- |
|
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" |
