roshan-soni/indicCorpV2-500k

Indic Unified Corpus (8M)

A 8.25M-row multilingual Indic text corpus spanning 21 languages, built by streaming ai4bharat/IndicCorpV2 and unifying it into a single dataset with a lang tag column.

Designed as monolingual pretraining data for multilingual encoder-decoder (denoising / translation) models on Indic languages.

Dataset Details

• Rows: 8,283,550
• Languages: 22
• Columns: text (string), lang (string)
• Size: ~1.56 GB (Arrow)
• Source: ai4bharat/IndicCorpV2
• License: CC-BY-4.0 (inherited from source)
• Language tag: ISO 639-1/639-3 code matching the source config

Per-language row counts

Language	Code	Rows
Assamese	asm	500,000
Bengali	ben	500,000
Bodo	brx	84,534
Dogri	doi	6,017
Konkani	gom	500,000
Gujarati	guj	500,000
Kannada	kan	500,000
Kashmiri	kas	3,221
Maithili	mai	426,657
Malayalam	mal	500,000
Marathi	mar	500,000
Manipuri	mni	42,703
Nepali	nep	500,000
Odia	ori	500,000
Punjabi	pan	500,000
Sanskrit	san	500,000
Santali	sat	110,914
Sindhi	snd	109,504
Tamil	tam	500,000
Telugu	tel	500,000
Urdu	urd	500,000
Hindi	hin	500,000
Total	—	8,283,550

How to Use

Load with datasets

from datasets import load_dataset

ds = load_dataset("roshan-soni/indicCorpV2-500k", split="train")
print(ds)

Filter by language

tamil = ds.filter(lambda r: r["lang"] == "tam")
bengali = ds.filter(lambda r: r["lang"] == "ben")

Stream (no full download)

ds = load_dataset("roshan-soni/indicCorpV2-500k", split="train", streaming=True)
for row in ds:
    print(row["lang"], row["text"][:80])
    break

Temperature-balanced sampling

For low-resource Indic languages, sample with p_i ∝ n_i^τ where τ ∈ [0.3, 0.7]:

import numpy as np
from collections import Counter

counts = Counter(ds["lang"])
langs = list(counts)
p = np.array([counts[l] for l in langs], dtype=np.float64) ** 0.5
p /= p.sum()
sample_lang = np.random.choice(langs, p=p)

Licensing

Inherited from IndicCorpV2: CC-BY-4.0. See the IndicCorpV2 dataset card for full attribution.

Citation

@inproceedings{doddapaneni-etal-2022-towards,
    title = "Towards Leaving No {I}ndic Language Behind: Building Monolingual Corpora, Benchmark and Models for {I}ndic Languages",
    author = "Doddapaneni, Sumanth and Aralikatte, Rahul and Ramesh, Gowtham and Goyal, Shreya and Khapra, Mitesh M. and others",
    booktitle = "Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)",
    month = may,
    year = "2022",
    address = "Dublin, Ireland",
    publisher = "Association for Computational Linguistics",
    pages = "6930--6946",
    url = "https://aclanthology.org/2022.acl-long.478",
    doi = "10.18653/v1/2022.acl-long.478",
}

Built from: ai4bharat/IndicCorpV2 · Last updated: 2026-09-11