metadata
dataset_info:
features:
- name: input_ids
sequence: int32
splits:
- name: train
num_bytes: 35168157552
num_examples: 1073116
download_size: 18303136476
dataset_size: 35168157552
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 dataset using the gemma 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, 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"