Aananda-giri/nepali_llm_datasets

Nepali LLM Datasets

This repository contains two configurations of Nepali LLM datasets:

Configurations

1. Scrapy Engine

• Description: Contains data collected using a web scraping engine.
• Files: [List any specific files or formats]

2. Nepberta

• Description: This dataset is derived from the Nepberta project and contains cleaned data specifically related to the project. The dataset contains **cleaned text chunks of size ~50 mb ** of all articles into a single giant string, with each article ending in <|endoftext|>. This long string is then segmented into chunks, each approximately 500 MB in size.
• Files: contains 23 files each ~500Mb (chunk_1.txt, chunk_2.txt, ... chunk_23.txt)
• split:train
  • files: chunk_1.txt to chunk_18.txt
• split:test
  • files: chunk_19.txt to chunk_23.txt

Usage

To load the datasets:

# it loads entire dataset first
from datasets import load_dataset

# Load nepberta configuration
nepberta_dataset = load_dataset("Aananda-giri/nepali_llm_datasets", name="nepberta", split='train') # use `streaming=True` to avoid downloading all the dataset

# length of chunks
len(nepberta_train['text']) # 18 : number of chunks
len(nepberta_train['text'][0])  # length of large text equivalent to 500 MB text

# use streaming=True to avoid downloading entire dataset
nepberta_train = load_dataset("Aananda-giri/nepali_llm_datasets", name="nepberta", streaming=True)['train']

# using next
next_text_chunk = next(iter(nepberta_train))
print(len(next_text_chunk['text']))

# using for loop
for large_chunk in nepberta_train:
  print(len(large_chunk['text']))
  break

  # code to process large_chunk['text']

# Load scrapy engine data
scrapy_train = load_dataset("Aananda-giri/nepali_llm_datasets", name="scrapy_engine" split="train")

pre-tokenized

IRIISNEPAL_U_Nepberta

• these files use context len. 512 and stride 384 (.75 * context_length)
• pre_tokenized/iriisnepal_u_nepberta_test_512.parquet
• pre_tokenized/iriisnepal_u_nepberta_train_512.parquet