metadata
dataset_info:
features:
- name: input_ids
sequence: int32
- name: attention_mask
sequence: int8
- name: token_type_ids
sequence: int8
- name: labels
sequence: int64
- name: next_sentence_labels
dtype: int64
splits:
- name: train
num_bytes: 12337588320
num_examples: 1715460
- name: validation
num_bytes: 26423408
num_examples: 3674
download_size: 1076906240
dataset_size: 12364011728
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
- split: validation
path: data/validation-*
Preprocessed dataset intended for pretraining of base models. Dataset was inspired by BERT pretraining - it includes NSP and MLM tasks and labels.