README.md

---
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: language_script
    dtype: string
  - name: minhash_cluster_size
    dtype: int64
  - name: top_langs
    dtype: string
  splits:
  - name: train
    num_bytes: 13252144546
    num_examples: 2293647
  download_size: 6366393037
  dataset_size: 13252144546
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
---
## Small Arabic FineWeb2 Sample Dataset  
A small extract (2.3 million rows compared to 58 million in the original: 
FineWeb2 [arb_Arab subset](https://huggingface.co/datasets/HuggingFaceFW/fineweb-2/viewer/arb_Arab)). 
First, I filtered for items with 95% or more Arabic (language_score is not reliable), 
then I randomly sampled the 2.3M from the result. 
See [this post](https://www.linkedin.com/posts/akhooli_a-small-arabic-fineweb2-sample-dataset-activity-7283099806003060736-g5cq)  
Code: 
```python
from datasets import load_dataset 
import pandas as pd 
from pprint import pprint
ds = load_dataset("akhooli/fineweb2_ar_24_sample")
import random
max_n = len(ds['train'])
index = random.randint(0,max_n)
pprint(ds['train'][index]['text'])
```