BhashaKritika: Building Synthetic Pretraining Data at Scale for Indic Languages

• You can find the paper on BhashaKritika here : Paper

1. Introduction

BhashaKritika is a large-scale synthetic pretraining corpus for 10 Indic languages. It is built using five generation strategies, including document-grounded, persona-based, topic-guided, and translation-based approaches.

The dataset is part of a systematic study on how grounding, instruction language, and native vs. translated generation affect data quality in multilingual settings. To ensure consistency at scale, we develop a modular quality evaluation pipeline with script and language detection, metadata checks, n-gram repetition analysis, and KenLM-based perplexity filtering.

BhashaKritika aims to provide a reliable, diverse, and linguistically rich synthetic corpus for pretraining high-quality Indic LLMs.

---

2. Dataset Details

Each entry captures both the generated text and detailed quality evaluation metadata.

Each sample includes:

Core Metadata

• master_id — Unique identifier for each generated instance.
• generation_technique — Method used for text generation (e.g., document_grounded, persona_based, topic_based, math_and_reasoning_based, translation_based)
• source — Origin of the context used for generation (indic_cc, fineweb2, etc.)
• style — Output style or format of the generation
• language — Target Indic language
• topic — Topical domain of the sample
• prompt — Input instruction provided to the model
• response — Generated output text

Quality & Safety Metadata

• flags — Indicators for automatically detected quality issues
• language_detection — Language identified
• word_count — Word-count of the generated text
• word_n_gram_repetition — 6-gram repetition scores
• perplexity — KenLM-based fluency/naturalness score
• quality_classification — Quality label and score assigned by the Fasttext Quality Classifier
• nsfw_words — Detected sensitive or inappropriate words ratio
• stop_words — Stopword occurrences ratio based on language-specific lists
• non_li_words — Words outside latin and indic alphabets

This structure provides a rich combination of generation details, linguistic analysis, and quality signals, enabling fine-grained filtering and large-scale pretraining for multilingual Indic LLMs.

---

3. How to Use and Run

You can load the dataset using the datasets library:

from datasets import load_dataset

ds = load_dataset(
    "krutrim-ai-labs/BhashaKritika",
    name="bengali",
    split="train"
)

---

4. License

This repository is licensed under the Krutrim Community License.

5. Citation

@misc{manoj2025bhashakritikabuildingsyntheticpretraining,
      title={BhashaKritika: Building Synthetic Pretraining Data at Scale for Indic Languages}, 
      author={Guduru Manoj and Neel Prabhanjan Rachamalla and Ashish Kulkarni and Gautam Rajeev and Jay Piplodiya and Arul Menezes and Shaharukh Khan and Souvik Rana and Manya Sah and Chandra Khatri and Shubham Agarwal},
      year={2025},
      eprint={2511.10338},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2511.10338}, 
}