Datasets:

agentlans
/

finefineweb-equal-weighted

Modalities:

Tabular

Text

Formats:

Languages:

Size:

Tags:

Libraries:

License:

Dataset card Data Studio Files Files and versions

xet

Community

finefineweb-equal-weighted / README.md

agentlans

Upload 3 files

753f783 verified 7 days ago

preview code

Raw

History Blame Contribute Delete

5.71 kB

metadata

license: apache-2.0
language:
  - en
tags:
  - text-classification
  - web-corpus
  - balanced-dataset
size_categories:
  - 100k<n<1M

Unofficial m-a-p/FineFineWeb Equal Weighted Sample

A balanced, class-weighted subset of m-a-p/FineFineWeb optimized for multi-class text classification tasks.

📊 Dataset Overview

To prevent class imbalance issues during model training, this dataset extracts an equal-weighted sample across all fine-grained domain classes.

Original Dataset: m-a-p/FineFineWeb
Total Samples: 134,000 (0 duplicates)
Training Set (85%): 113,900 rows
Validation Set (15%): 20,100 rows
Classes: 67 balanced domains (2,000 samples per class)

🚀 Quick Start

Load the dataset easily using the Hugging Face datasets library:

from datasets import load_dataset

dataset = load_dataset("agentlans/finefineweb-equal-weighted")
print(dataset["train"][0])

📂 Data Structure

Each row contains text metadata, source tracking information, and the assigned domain label (domain).

Example Row

{
  "url": "https://lothbury.exchange/parson-desires-cure-fund-for-local-utilities/",
  "date": "2023-03-27T13:42:03Z",
  "file_path": "s3://commoncrawl/crawl-data/CC-MAIN-2023-14/segments/1679296948632.20/warc/CC-MAIN-20230327123514-20230327153514-00428.warc.gz",
  "language_score": 0.93825,
  "token_count": 894,
  "dump": "CC-MAIN-2023-14",
  "global_id": "webtext-fineweb__CC-MAIN-2023-14__0__10195829",
  "lang": "en",
  "text": "Mike Parson is actually asking for an emergency therapy account...",
  "domain": "journalism_and_media_communication",
  "round": 1,
  "source": "journalism_and_media_communication/journalism_and_media_communication_001222.jsonl"
}

Note: The source field indicates the exact origin file from which the row was pulled.

📋 Class Distribution

All 67 categories maintain a strict 1:1 balance of 1,700 training samples and 300 validation samples (2,000 total per class).

Label	Train	Validation	Total
aerospace	1700	300	2000
agronomy	1700	300	2000
artistic	1700	300	2000
astronomy	1700	300	2000
atmospheric_science	1700	300	2000
automotive	1700	300	2000
beauty	1700	300	2000
biology	1700	300	2000
celebrity	1700	300	2000
chemistry	1700	300	2000
christianity	1700	300	2000
civil_engineering	1700	300	2000
communication_engineering	1700	300	2000
computer_science_and_technology	1700	300	2000
design	1700	300	2000
drama_and_film	1700	300	2000
economics	1700	300	2000
electronic_science	1700	300	2000
entertainment	1700	300	2000
environmental_science	1700	300	2000
fashion	1700	300	2000
finance	1700	300	2000
food	1700	300	2000
gamble	1700	300	2000
game	1700	300	2000
geography	1700	300	2000
health	1700	300	2000
history	1700	300	2000
hobby	1700	300	2000
hydraulic_engineering	1700	300	2000
instrument_science	1700	300	2000
journalism_and_media_communication	1700	300	2000
landscape_architecture	1700	300	2000
law	1700	300	2000
library	1700	300	2000
literature	1700	300	2000
materials_science	1700	300	2000
mathematics	1700	300	2000
mechanical_engineering	1700	300	2000
medical	1700	300	2000
mining_engineering	1700	300	2000
movie	1700	300	2000
music_and_dance	1700	300	2000
news	1700	300	2000
nuclear_science	1700	300	2000
ocean_science	1700	300	2000
optical_engineering	1700	300	2000
painting	1700	300	2000
pet	1700	300	2000
petroleum_and_natural_gas_engineering	1700	300	2000
philosophy	1700	300	2000
photo	1700	300	2000
physics	1700	300	2000
politics	1700	300	2000
psychology	1700	300	2000
public_administration	1700	300	2000
relationship	1700	300	2000
sociology	1700	300	2000
sports	1700	300	2000
statistics	1700	300	2000
systems_science	1700	300	2000
textile_science	1700	300	2000
topicality	1700	300	2000
transportation_engineering	1700	300	2000
travel	1700	300	2000
urban_planning	1700	300	2000
weapons_science	1700	300	2000
Total	113900	20100	134000

📝 License & Citation

This sample is distributed under the Apache 2.0 License.

If you use this balanced dataset in your research or applications, please cite the original FineFineWeb authors:

@misc{map2024finefineweb,
  title={FineFineWeb: A Comprehensive Study on Fine-grained Domain Web Corpus},
  url={https://huggingface.co/datasets/m-a-p/FineFineWeb},
  author={Zhang, Ge and Du, Xinrun and Yu, Zhimiao and Wang, Zili and Wang, Zekun and Guo, Shuyue and Zheng, Tianyu and Zhu, Kang and Liu, Jerry and Yue, Shawn and Liu, Binbin and Peng, Zhongyuan and Yao, Yifan and Yang, Jack and Li, Ziming and Zhang, Bingni and Liu, Minghao and Liu, Tianyu and Gao, Yang and Chen, Wenhu and Zhou, Xiaohuan and Liu, Qian and Wang, Taifeng and Huang, Wenhao},
  publisher={Hugging Face},
  version={v0.1.0},
  month={December},
  year={2024}
}