README.md

metadata

license: cc-by-4.0
task_categories:
  - text-generation
language:
  - as
  - bn
  - gu
  - en
  - hi
  - kn
  - ks
  - ml
  - mr
  - ne
  - or
  - pa
  - sa
  - sd
  - ta
  - te
  - ur
tags:
  - language-modeling
  - casual-lm
  - llm
pretty_name: sangraha
dataset_info:
  - config_name: verified
    features:
      - name: doc_id
        dtype: string
      - name: type
        dtype: string
      - name: text
        dtype: string
    splits:
      - name: asm
      - name: ben
      - name: brx
      - name: doi
      - name: eng
      - name: gom
      - name: guj
      - name: hin
      - name: kan
      - name: kas
      - name: mai
      - name: mal
      - name: mar
      - name: mni
      - name: nep
      - name: ori
      - name: pan
      - name: san
      - name: sat
      - name: snd
      - name: tam
      - name: tel
      - name: urd
  - config_name: unverified
    features:
      - name: doc_id
        dtype: string
      - name: text
        dtype: string
    splits:
      - name: asm
      - name: ben
      - name: guj
      - name: hin
      - name: kan
      - name: mal
      - name: mar
      - name: nep
      - name: ori
      - name: pan
      - name: san
      - name: tam
      - name: tel
      - name: urd
configs:
  - config_name: verified
    data_files:
      - split: asm
        path: verified/asm.*
      - split: ben
        path: verified/ben.*
      - split: brx
        path: verified/brx.*
      - split: doi
        path: verified/doi.*
      - split: eng
        path: verified/eng.*
      - split: gom
        path: verified/gom.*
      - split: guj
        path: verified/guj.*
      - split: hin
        path: verified/hin.*
      - split: kan
        path: verified/kan.*
      - split: kas
        path: verified/kas.*
      - split: mai
        path: verified/mai.*
      - split: mal
        path: verified/mal.*
      - split: mar
        path: verified/mar.*
      - split: mni
        path: verified/mni.*
      - split: nep
        path: verified/nep.*
      - split: ori
        path: verified/ori.*
      - split: pan
        path: verified/pan.*
      - split: san
        path: verified/san.*
      - split: sat
        path: verified/sat.*
      - split: snd
        path: verified/snd.*
      - split: tam
        path: verified/tam.*
      - split: tel
        path: verified/tel.*
      - split: urd
        path: verified/urd.*
  - config_name: unverified
    data_files:
      - split: asm
        path: unverified/asm.*
      - split: ben
        path: unverified/ben.*
      - split: guj
        path: unverified/guj.*
      - split: hin
        path: unverified/hin.*
      - split: kan
        path: unverified/kan.*
      - split: mal
        path: unverified/mal.*
      - split: mar
        path: unverified/mar.*
      - split: nep
        path: unverified/nep.*
      - split: ori
        path: unverified/ori.*
      - split: pan
        path: unverified/pan.*
      - split: san
        path: unverified/san.*
      - split: tam
        path: unverified/tam.*
      - split: tel
        path: unverified/tel.*
      - split: urd
        path: unverified/urd.*

Sangraha

Sangraha is the largest Indic language pretraining data containing 251B tokens summed up over 22 languages, extracted from curated sources, existing multilingual corpora and large scale translations.

Coming Soon:

• Sangraha Synthetic
• Sangraha Verified - Hindi YouTube data

More information:

• Read more about Sangraha on Arxiv;
• Check out the scraping and cleaning pipelines used to curate Sangraha on GitHub;

Getting Started

You can download the dataset using Hugging Face datasets:

from datasets import load_dataset
ds = load_dataset("cerebras/SlimPajama-627B")

Background

Sangraha contains three broad components:

• Sangraha Verified: Containing scraped data from "human-verified" Websites, OCR-extracted data from high quality Indic language PDFs, transcribed data from various Indic language videos, podcasts, movies, courses, etc.
• Sangraha Unverfied: High quality Indic language data extracted from existing multilingual corpora employing perplexity filtering using n-gram language models trained on Sangraha Verified.
• Sangraha Synthetic: WikiMedia English translated to 14 Indic languages and further "romanised" from 14 languages by transliteration to English.