raj2708/fineweb-filtered

FineWeb Filtered

A filtered subset of HuggingFaceFW/fineweb.
Only high-quality, clean English web documents are kept. No text is modified — filtering is purely keep/drop.

---

What's in this dataset

Each record is a web document with these columns (inherited from FineWeb):

Column	Description
text	Full document text
id	Unique document ID
url	Source URL
dump	Common Crawl dump (e.g. CC-MAIN-2024-10)
date	Crawl date
language	Detected language code
language_score	Language-ID confidence score
token_count	Estimated token count

---

Filters applied

Documents are dropped if they fail any of these checks (in order):

Filter	Condition to DROP
quality	Less than 200 characters or 50 words · language score below 0.65 · symbol ratio above 30% · digit ratio above 40%
license	Matches a blocked license or domain (default: none blocked)
ai_generated	AI-probability above 0.85
toxicity	Toxicity score above 0.5
perplexity	Perplexity outside the normal range (uses entropy proxy if no KenLM model is set)

---

How to use

Load with 🤗 Datasets

from datasets import load_dataset

ds = load_dataset("raj2708/fineweb-filtered", split="train", streaming=True)

for row in ds:
    print(row["text"])
    break

Load a single Parquet file directly

import pandas as pd

df = pd.read_parquet(
    "hf://datasets/raj2708/fineweb-filtered/train/part-000001.parquet"
)
print(df.head())

Filter by language score in code

ds = load_dataset("raj2708/fineweb-filtered", split="train", streaming=True)
high_conf = ds.filter(lambda x: x["language_score"] > 0.9)

---

Source & License

Built from HuggingFaceFW/fineweb — licensed ODC-By 1.0.
This dataset inherits the same license. Please cite the original FineWeb when using this data.