metadata
dataset_info:
features:
- name: input_ids
sequence: int32
- name: attention_mask
sequence: int8
splits:
- name: train
num_bytes: 3977615851
num_examples: 2293647
download_size: 1879839994
dataset_size: 3977615851
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
This dataset is an Arabic sample extracted from the Fineeb2
Arabic subset (arb_Arab) which is supposed to be standard Arabic.
There are around 2.3 million rows in this sample. First, the whole dataset (57.8M rows) was scanned and rows
were kept if they had over 95% Arabic words. Then this 2.3M sample was randomly sampled from the mostly Arabic data.
Notice that language_score is not an accurate measure. Also, this did not exclude slang, dialects or inappropriate
content (no editing was done to any row and all columns were kept).
The main purpose of this dataset is educational and I hope it helps researchers in designing and developing pre-processing
for the main FineWeb2 dataset (or any other Arabic corpora).
Example:
from datasets import load_dataset
from pprint import pprint
import random
ds = load_dataset("akhooli/fineweb2_ar_24_sample")
max_n = len(ds['train'])
index = random.randint(0,max_n) # random row
pprint(ds['train'][index]['text']) # article