Dataset Card for eoinf/pile_llama_800m

Original Dataset

Original dataset: monology/pile-uncopyrighted

Dataset Details

• Total Tokens: 816,116,736
• Total Sequences: 796,989
• Context Length: 1024 tokens
• Tokenizer: meta-llama/Llama-2-7b-hf
• Format: Each example contains a single field tokens with a list of 1024 token IDs

Preprocessing

Each document was:

1. Tokenized using the meta-llama/Llama-2-7b-hf tokenizer
2. Prefixed with a BOS (beginning of sequence) token
3. Suffixed with an EOS (end of sequence) token
4. Packed into fixed-length sequences of 1024 tokens

Usage

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("eoinf/pile_llama_800m")

# Access training data
train_data = dataset["train"]
print(train_data[0]["tokens"])  # First sequence

Use with PyTorch

import torch
from datasets import load_dataset
from torch.utils.data import DataLoader

dataset = load_dataset("eoinf/pile_llama_800m", split="train")

# Convert to PyTorch tensors
dataset.set_format(type="torch", columns=["tokens"])

# Create DataLoader
dataloader = DataLoader(dataset, batch_size=32, shuffle=True)

for batch in dataloader:
    tokens = batch["tokens"]  # Shape: (batch_size, 1024)
    # Your training code here

Dataset Statistics

• Total storage size: ~1.63 GB (as uint16)
• Sequences per batch folder: 10,000
• Number of batch folders: 80