hemantvirmani/gpt-training-dataset

📚 GPT Training Dataset (WikiText + OpenWebText Mix)

Overview

This dataset is a cleaned and curated text corpus designed for training small to mid-sized GPT-style language models.

It combines:

• WikiText-103 (high-quality structured text)
• OpenWebText (real-world web text, sampled)

The goal is to provide a balanced dataset that: - trains quickly - produces coherent text - avoids excessive noise from large web corpora

Dataset Composition

The final corpus is a mixture of high-quality encyclopedic text and diverse web content, designed to balance factual density with natural conversational flow.

Source	Proportion	Description
WikiText-103	~75%	High-quality, verified articles from Wikipedia. Provides structured knowledge.
OpenWebText	~25%	Sampled web content filtered for quality. Provides stylistic variety.

Data Splits

The dataset is provided as a single dataset.txt file, intended to be split as follows:

• Training (80%): Primary data for model weight updates.
• Validation (20%): Used for calculating perplexity and monitoring over-fitting during training.

Total size: ~1.33 GB (uncompressed text).

Technical Specifications

Raw Text: dataset.txt (1.33 GB) Tokenized Data (binary): dataset.bin (574 MB) — Contiguous token IDs for training *Tokenized Data (numpy): tokens.npy (574 MB) — Same tokens as dataset.bin, stored as a NumPy array

---

Preprocessing Pipeline

The dataset was generated using a custom script with the following steps:

1. Cleaning

• Removed section headers (e.g., == Title ==)
• Normalized spacing and punctuation artifacts
• Stripped malformed tokens

2. Filtering

• Removed very short or low-quality lines
• Ensured minimum text length and structure

3. Deduplication

• Removed duplicate entries (keeping one copy per unique sample)

4. Shuffling

• Randomized dataset order for better training distribution

5. Document Separation

• Each sample is separated using:

<|endoftext|>

---

Files

dataset.txt

• Cleaned text dataset
• One document per <|endoftext|> separator
• Recommended for:
  • custom tokenization
  • experimentation

---

dataset.bin

• Pre-tokenized binary file
• Format: uint16
• Tokenizer: GPT-2 (tiktoken)

Load example:

import numpy as np

data = np.memmap("dataset.bin", dtype=np.uint16, mode="r")

---

tokens.npy

• Same tokenized data in NumPy format
• Useful for debugging and inspection

Load example:

import numpy as np

tokens = np.load("tokens.npy", mmap_mode="r")

---

Tokenization

Pre-tokenization (for .bin and .npy) uses:

• GPT-2 tokenizer via tiktoken

import tiktoken
enc = tiktoken.get_encoding("gpt2")

⚠️ Important:
If using dataset.bin, you must use the same tokenizer for inference.

---

Usage (PyTorch Example)

import numpy as np
import torch

data = np.memmap("dataset.bin", dtype=np.uint16, mode="r")
data = torch.from_numpy(data.astype(np.int64))

block_size = 128
batch_size = 32

def get_batch():
    ix = torch.randint(len(data) - block_size, (batch_size,))
    x = torch.stack([data[i:i+block_size] for i in ix])
    y = torch.stack([data[i+1:i+block_size+1] for i in ix])
    return x, y

---

Intended Use

This dataset is ideal for:

• training GPT-style models from scratch
• experimentation with small architectures
• educational purposes and learning pipelines

---

Limitations

• Not suitable for large-scale production LLM training
• Limited domain diversity compared to massive corpora
• OpenWebText portion may still contain minor noise

---

Reproducibility

Dataset can be regenerated using the provided script:

python prepare_dataset.py

---

Acknowledgements

• WikiText-103
• OpenWebText

---

License

Please refer to the original dataset licenses for:

• WikiText-103
• OpenWebText

---

Author

Hemant Virmani created this as part of a GPT training pipeline experiment.