metadata

pretty_name: edu_fineweb10B_sharded_50shards
language:
  - en
tags:
  - language-modeling
  - tokenized
  - tiktoken
  - npy
  - gpt2
  - sharded
  - autoregressive
task_categories:
  - text-generation
size_categories:
  - 10B+
license: mit

Dataset Card for edu_fineweb10B_sharded_50shards

This dataset card aims to describe the edu_fineweb10B_sharded_50shards dataset, a large-scale pre-tokenized and sharded dataset created from the eduFineWeb corpus. It has been prepared for use in training transformer-based language models using NumPy arrays for efficient loading.

Dataset Details

Dataset Description

edu_fineweb10B_sharded_50shards is a tokenized dataset based on the eduFineWeb 10B corpus, designed for scalable training of language models. The dataset contains a total of 10 billion tokens split across 50 shards—49 for training and 1 for evaluation.

Each shard is stored in .npy format and contains 200 million token IDs, pre-tokenized using the tiktoken tokenizer (compatible with GPT-2-style models). The dataset is designed for high-efficiency training pipelines, particularly in distributed or multi-GPU setups.

Curated by: Private (individual researcher)
Funded by [optional]: Not funded
Shared by [optional]: Abhinav
Language(s) (NLP): English
License: MIT

Dataset Sources

Repository: [Private / Local Project]
Paper [optional]: N/A
Demo [optional]: N/A

Uses

Direct Use

The dataset is suitable for:

Pretraining large autoregressive language models (e.g., GPT-style)
Finetuning models for general-purpose language generation
Research in scaling laws, memory-efficient training, and model evaluation

Out-of-Scope Use

Not suited for supervised learning tasks without additional labeling
Not appropriate for real-time applications without additional safety filtering
Not recommended for use in sensitive, safety-critical environments without thorough auditing

Dataset Structure

Format: .npy files
Tokenizer used: tiktoken.get_encoding("gpt2")
Tokens per shard: 200 million
Total shards: 50
Shard naming format:
- Training: edu_fineweb_train_000000.npy to edu_fineweb_train_000048.npy
- Evaluation: edu_fineweb_val_000000.npy
File size: ~~400MB per shard (~~20GB total)

Each file contains a flat NumPy array of token IDs (int32), which can be converted to PyTorch tensors for training.

Example Usage

import numpy as np
import torch

# Load a shard file
filename = "edu_fineweb_train_000001.npy"
npt = np.load(filename)
npt = npt.astype(np.int32)
ptt = torch.tensor(npt, dtype=torch.long)

print(ptt.shape)  # Should be (200_000_000,)
print(ptt[:10])   # View first 10 token IDs