README.md

metadata

dataset_info:
  features:
    - name: text
      dtype: string
    - name: id
      dtype: string
    - name: dump
      dtype: string
    - name: url
      dtype: string
    - name: date
      dtype: string
    - name: file_path
      dtype: string
    - name: language
      dtype: string
    - name: language_score
      dtype: float64
    - name: token_count
      dtype: int64
    - name: score
      dtype: float64
    - name: int_score
      dtype: int64
    - name: token_ids
      sequence: int64
  splits:
    - name: train
      num_bytes: 83423842611
      num_examples: 2494618
  download_size: 32521124201
  dataset_size: 83423842611
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
task_categories:
  - text-generation
language:
  - en
splits:
  - name: train
    num_bytes: 41746813043
    num_examples: 2494618
    download_size: 9359508369
    dataset_size: 41746813043

Processed FineWeb-Edu Dataset

Dataset Name on Hugging Face: PursuitOfDataScience/processed-fineweb-edu

Overview

This dataset is a processed version of the FineWeb-Edu dataset, intended for language model training and NLP research. It has been tokenized and truncated according to a specified block size (i.e., 2048), preparing it for model pre-training or evaluation with transformer-based language models.

Source Dataset

• Name: FineWeb-Edu
• Description: A dataset focused on educational text extracted from the web, designed for language modeling and educational NLP tasks.
• Link: https://huggingface.co/datasets/HuggingFaceFW/fineweb-edu
• Version: CC-MAIN-2024-10

Processing Steps

The dataset was processed using the Hugging Face Datasets library and a Hugging Face tokenizer. The primary steps include:

1. Tokenization: Each text sample is encoded using the tokenizer’s .encode() method.
2. Truncation: Token sequences are truncated to a specified block_size + 1.
3. Filtering: Any sample with fewer than block_size + 1 tokens is removed.
4. Saving: The processed data is saved to disk using ds.save_to_disk(processed_dir).

Below is the code excerpt used to perform these steps:

def load_nonstream_data(data_files, hf_tokenizer, block_size, num_proc=128):
    """
    Loads the entire dataset in memory either from a cached processed directory
    or processes it in parallel if not yet cached.
    Returns a list of token ID sequences.
    """

    processed_dir = "processed_data/tokenized_data"
    if os.path.exists(processed_dir):
        print(f"Loading cached dataset from '{processed_dir}'...")
        ds = load_from_disk(processed_dir)
        tokenized_data = ds["token_ids"]
        return tokenized_data

    print("No cached dataset found. Processing in parallel...")

    ds_dict = load_dataset("arrow", data_files=data_files, streaming=False)
    if "train" in ds_dict:
        ds = ds_dict["train"]
    else:
        ds = ds_dict

    def tokenize_and_truncate(example):
        text = example["text"] if "text" in example else ""
        token_ids = hf_tokenizer.encode(text)
        if len(token_ids) < block_size + 1:
            return {"token_ids": None}
        token_ids = token_ids[:block_size+1]
        return {"token_ids": token_ids}

    ds = ds.map(
        tokenize_and_truncate,
        batched=False,
        num_proc=num_proc
    )
    ds = ds.filter(lambda ex: ex["token_ids"] is not None, num_proc=num_proc)

    if "text" in ds.column_names:
        ds = ds.remove_columns(["text"])

    os.makedirs(os.path.dirname(processed_dir), exist_ok=True)
    ds.save_to_disk(processed_dir)
    print(f"Processed dataset saved to '{processed_dir}'.")

    tokenized_data = ds["token_ids"]
    return tokenized_data

Dataset Structure

• Columns:

  • token_ids: A list of token IDs representing a truncated text segment.

• Splits:

  • This dataset is provided as a single split named train.

Intended Use & Applications

• Language Modeling: Suitable for GPT-style or other auto-regressive models, focusing on educational text.
• Fine-Tuning: Can be used to fine-tune existing models on educational text.
• Research: Useful for experimentation in NLP tasks such as text generation.

How to Load

You can load this dataset directly from Hugging Face using the datasets library:

from datasets import load_dataset

dataset = load_dataset("PursuitOfDataScience/processed-fineweb-edu")
print(dataset)