---
dataset_info:
  features:
  - name: content
    dtype: string
  - name: score
    dtype: float64
  - name: source
    dtype: string
  splits:
  - name: train
    num_bytes: 404127039
    num_examples: 100060
  download_size: 233844039
  dataset_size: 404127039
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
---

### Sampling Code

```py
from __future__ import annotations

import os
from typing import List

from datasets import Dataset, load_dataset
from tqdm.auto import tqdm

TARGET_CHARS = 400_000_000  # ≈100 M tokens (≈4 chars/token)
BUFFER = 10_000             # streaming shuffle buffer
SEED = 42
HF_REPO = "sumuks/Ultra-FineWeb-100M"


def sample_ultrafineweb(
    target_chars: int = TARGET_CHARS,
    buffer_size: int = BUFFER,
    seed: int = SEED,
) -> Dataset:
    """
    Stream the Ultra‑FineWeb English split and return a random sample
    whose total `content` length is at least `target_chars` characters
    (≈100 M tokens).
    """
    stream = load_dataset("openbmb/Ultra-FineWeb", split="en", streaming=True)
    stream = stream.shuffle(seed=seed, buffer_size=buffer_size)

    picked: List[dict] = []
    char_count = 0

    for row in tqdm(stream, desc="sampling"):
        text = row["content"]
        char_count += len(text)
        picked.append(row)
        if char_count >= target_chars:
            break

    return Dataset.from_list(picked)


def main() -> None:
    ds = sample_ultrafineweb()
    total_chars = sum(len(r["content"]) for r in ds)
    print(f"Sampled {len(ds):,} documents, {total_chars:,} characters")
    ds.push_to_hub(HF_REPO, private=False)


if __name__ == "__main__":
    main()

```