Research map / Large language models / Language, safety and interpretability
BhashaKritika: Building Synthetic Pretraining Data at Scale for Indic Languages
Guduru Manoj, Neel Prabhanjan Rachamalla, Ashish Kulkarni, Gautam Rajeev, Jay Piplodiya, Arul Menezes, Shaharukh Khan, Souvik Rana, Manya Sah, Chandra Khatri, Shubham Agarwal
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use synthetic BhashaKritika Indic data instead of Web data in annealing | LLaMA-3.2 1B · Indic Sentiment | accuracy: 0.617 -> 0.631 | Improved |
| use synthetic data instead of Web in annealing | LLaMA-3.2 1B · GSM8K | accuracy: 0.111 -> 0.120 | Improved |
| use synthetic data instead of Web in annealing | LLaMA-3.2 1B · Indic XParaphrase | accuracy: 0.782 -> 0.736 | Hurt |
| use synthetic data instead of Web in annealing | LLaMA-3.2 1B · Indic XNLI | accuracy: 0.433 -> 0.403 | Hurt |
| use synthetic data instead of Web in annealing | LLaMA-3.2 1B · Hellaswag, MMLU, OpenbookQA | accuracy: 0.483 -> 0.482 (Hellaswag); 0.412 -> 0.408 (MMLU) | No clear effect |
| continue pretraining on synthetic data instead of repeated Web epochs | LLaMA-3.2 1B from scratch · Indic XNLI | accuracy: 0.330 -> 0.339 | Improved |
| continue pretraining on synthetic data instead of repeated Web epochs | LLaMA-3.2 1B from scratch · ARC Challenge Indic | accuracy: 0.213 -> 0.210 | No clear effect |
| add English prompt to Indic source document (vs Indic prompt) (ablation) | document-grounded generation · Gujarati discard rate | discard rate (%): 24.15 -> 15.85 | Improved |
| use Indic document instead of English document with English prompt (ablation) | document-grounded generation · Gujarati discard rate | discard rate (%): 17.45 -> 15.85 | Improved |
| use Indic prompt instead of English prompt for Indic documents (ablation) | document-grounded generation · Bengali discard rate | discard rate (%): 0.47 -> 0.33 | Improved |
| add matching document grounding to English persona (ablation) | persona-based generation · discard rate | discard rate (%): 1.93% -> 3.50% | Hurt |
| add random document grounding to English persona (ablation) | persona-based generation · discard rate | discard rate (%): 1.93% -> 14.43% | Hurt |
Closest papers by meaning
- Towards Scalable Data Diversification for Language Model Pretraining via Leverage Score Sampling (NeurIPS 2026 · similarity 0.88)
- Enhancing Multilingual LLM Pretraining with Model-Based Data Selection (NeurIPS 2025 · similarity 0.88)
- MuRating: A High Quality Data Selecting Approach to Multilingual Large Language Model Pretraining (NeurIPS 2025 · similarity 0.88)
- Towards a Theoretical Understanding of Synthetic Data in LLM Post-Training: A Reverse-Bottleneck Perspective (ICLR 2025 · similarity 0.88)
- How to Synthesize Text Data without Model Collapse? (ICML 2025 · similarity 0.87)
- RedPajama: an Open Dataset for Training Large Language Models (NeurIPS 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
