DaruLM dataset for LLM adaptation

Description

A growing collection of texts of various domains for Russian LLM adaptation extracted from other Hugging Face datasets and open resources.

Usage of this dataset is possible only for scientific purposes on a non-commercial basis.

Credits: Initial datasets were provided by Ilya Gusev

NOTICE: Some domain splits are based on vocabulary stats and may be noisy

Current domains: (used for domains argument in load_datasets):

accounting	antique	aphorisms	art
biography	biology	buriy	business
cinema	computers	design	dramaturgy
economics	enwiki	essay	fantasy
gazeta	geography	guidebooks	habr
history	humor	language	law
lenta	literature	medicine	military
music	ods-tass	philosophy	pikabu
politic	prose	psychology	reference
religion	science	sociology	taiga-fontanka
textbook	wiki	UNDEFINED

Usage

Prerequisites:

pip install datasets zstandard jsonlines pysimdjson

Dataset iteration:

import datasets

# Load habr and textbooks
for example in datasets.load_dataset('dichspace/darulm', domains=["habr","textbook"], split="train", streaming=True):
    print(example.keys())
    print(example)
    break