File size: 3,553 Bytes
229615b a94283d 229615b a94283d 229615b | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 | ---
language:
- en
license: openrail
size_categories:
- 10M<n<100M
task_categories:
- fill-mask
tags:
- bert
- mlm
- masked-language-modeling
- wikipedia
- bookcorpus
dataset_info:
features:
- name: text
dtype: string
splits:
- name: train
num_bytes: 22046503784.16388
num_examples: 72440303
- name: validation
num_bytes: 1224805681.247909
num_examples: 4024461
- name: test
num_bytes: 1224805985.5882123
num_examples: 4024462
download_size: 15416904536
dataset_size: 24496115451.0
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
- split: validation
path: data/validation-*
- split: test
path: data/test-*
---
# Unified English MLM Pre-training Corpus (80M Rows)
This dataset is a massive, diverse, multi-domain English text corpus explicitly engineered for pre-training and domain-adaptation of **BERT-style models via Masked Language Modeling (MLM)**. It aggregates over 80 million rows of text, completely stripped of auxiliary metadata, labels, and identifiers to expose purely raw text strings.
## Dataset Details
- **Repository ID:** `8Opt/bert-mlm-experiments-en`
- **Total Rows:** 80,489,226
- **Format:** Single-column (`text`) format optimized for fast, block-based tokenization pipelines.
### Data Splits
| Split | Number of Rows | Percentage |
| :--- | :--- | :--- |
| **Train** | 72,440,303 | ~90% |
| **Validation** | 4,024,461 | ~5% |
| **Test** | 4,024,462 | ~5% |
_Note:_ This is the [code](https://colab.research.google.com/drive/1141Gi5gwB8KwnE-CQHAkTRpWSKuNZV1y?usp=sharing) I used to create this dataset.
---
## Dataset Composition & Provenance
This corpus was programmatically compiled, cleaned, and standardized from the following source datasets:
1. **BookCorpus** (`rojagtap/bookcorpus`): High-quality narrative text from thousands of unpublished books, vital for capturing long-range context and literary style.
2. **Wikipedia** (`wikimedia/wikipedia` - split `20231101.en`): Comprehensive, factual, and well-structured encyclopedic knowledge covering an expansive array of topics.
3. **IMDb Movie Reviews** (`ajaykarthick/imdb-movie-reviews`): Rich conversational, highly descriptive, and emotionally expressive language patterns.
4. **STS (Semantic Textual Similarity) 2012-2015** (`mteb/sts12-sts` through `sts15-sts`): Highly nuanced paired sentence variations, flattened into individual rows to expose the model to standalone structured sentence syntaxes.
---
## Processing Methodology
- **Schema Unification:** All incoming dataset features (titles, bodies, labels, scores, URLs) were stripped out or merged.
- **Structural Flattening:** For datasets containing sentence pairs (like STS) or multiple text fields (like Title/Body), the elements were completely flattened into standalone, individual rows under the `text` column to preserve line-by-line semantic structure.
- **Randomized Splitting:** The final dataset dictionary was generated using a reproducible two-step deterministic shuffle and split (`seed=42`).
---
## Intended Use & Quickstart
This dataset is primarily intended for Masked Language Modeling (MLM). Because of its massive validation split (~4M rows), it is highly recommended to subset the validation split during training check-ins to prevent evaluation bottlenecking.
### How to Load
```python
from datasets import load_dataset
# Load the full dataset dictionary (Train, Validation, Test)
dataset = load_dataset("8Opt/bert-mlm-experiments-en")
# Quick print overview
print(dataset) |