metadata

configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
  - config_name: text-only
    data_files:
      - split: train
        path: text-only/train-*
dataset_info:
  - config_name: default
    features:
      - name: url
        dtype: string
      - name: text
        dtype: string
      - name: date
        dtype: string
      - name: metadata
        dtype: string
    splits:
      - name: train
        num_bytes: 4467051029
        num_examples: 1820241
    download_size: 1772035124
    dataset_size: 4467051029
  - config_name: text-only
    features:
      - name: text
        dtype: string
    splits:
      - name: train
        num_bytes: 2305854627
        num_examples: 1820241
    download_size: 1360869461
    dataset_size: 2305854627
license: odc-by
task_categories:
  - text-generation
size_categories:
  - 1M<n<10M
source_datasets: open-web-math/open-web-math

Dataset Card for "open-web-math-minhash"

An attempt at a "high quality sample" of open-web-math/open-web-math by aggressively applying minhash from text-dedup. The result is 1.82M rows down from the original 6M:

DatasetDict({
    train: Dataset({
        features: ['url', 'text', 'date', 'metadata'],
        num_rows: 1820241
    })
})

Usage

Unless you need the metadata, load the text-only config which is only 1.4 GB/5 shards:

from datasets import load_dataset

dataset_config = "text-only"
dataset = load_dataset("BEE-spoke-data/open-web-math-minhash", dataset_config)

making of

On a high-RAM colab TPU (40 cores)

from pathlib import Path
from tqdm.auto import tqdm

ds_name = "open-web-math/open-web-math"
dataset_config = "default"
data_split = 'train'
text_column = 'text'

out_dir = Path(f"output/minhash/{ds_short_name}/{data_split}")
!mkdir -p $out_dir


!python -m text_dedup.minhash \
  --path $ds_name \
  --name $dataset_config \
  --split $data_split \
  --cache_dir "./cache" \
  --output $out_dir \
  --column $text_column \
  --ngram 5 --threshold 0.5 \
  --hash_func xxh3 --hash_bits 16 --num_perm 64 \
  --batch_size 10000

print(f"output dir is:\n\t{out_dir}")
!ls $out_dir

Console:

Resolving data files: 100% 114/114 [00:11<00:00,  9.79it/s]
Fingerprinting... (num_proc=40): 100% 6315233/6315233 [15:27<00:00, 6806.11 examples/s]
Iterating MinHashes...: 100% 632/632 [05:37<00:00,  1.87it/s]
Clustering...: 100% 14/14 [01:13<00:00,  5.22s/it]
Finding clusters... (num_proc=40): 100% 6315233/6315233 [10:57<00:00, 9602.90 examples/s]
Filtering clusters... (num_proc=40): 100% 6315233/6315233 [03:53<00:00, 27069.61 examples/s]
Saving the dataset (33/33 shards): 100% 1820241/1820241 [07:07<00:00, 4260.38 examples/s]
[10/11/23 23:41:46] INFO     Loading                         :

citation

@misc{paster2023openwebmath,
      title={OpenWebMath: An Open Dataset of High-Quality Mathematical Web Text}, 
      author={Keiran Paster and Marco Dos Santos and Zhangir Azerbayev and Jimmy Ba},
      year={2023},
      eprint={2310.06786},
      archivePrefix={arXiv},
      primaryClass={cs.AI}
}