Research
We publish what we build.
5 papers across AACL, EMNLP, and ICDM. The problems we solve in production, we also advance the science behind.
Publications
Published work
Data Science Task Benchmarking with Context Engineering
Ram Mohan Rao Kadiyala, Hamza Farooq et al.
We introduce a benchmark to evaluate LLM-based data science agents using real-world interaction data. Three models are tested across multiple task categories and prompting strategies. Our analysis examines model sensitivity to ambiguity, data leakage, and temperature variation. Results reveal key performance differences and provide a foundation for advancing reliable data science agents.
Ram Mohan Rao Kadiyala, Siddhant Gupta, Jebish Purbey, Giulio Martini, Ali Shafique, Suman Debnath, Hamza Farooq
Improving Multilingual Capabilities with Cultural and Local Knowledge in Large Language Models
Ram Mohan Rao Kadiyala, Hamza Farooq et al.
We present a Hindi-English bilingual LLM achieving a 3% average improvement across both languages, outperforming larger models. Trained on a curated 485K instruction dataset, it enhances multilingual performance without architectural changes. Over 140 training runs showed that culturally informed fine-tuning boosts bilingual ability efficiently. All models, data, and code are released under MIT and Apache licenses.
Ram Mohan Rao Kadiyala, Siddartha Pullakhandam, Siddhant Gupta, Jebish Purbey, Drishti Sharma, Kanwal Mehreen, Muhammad Arham, Suman Debnath, Hamza Farooq
Alif: Advancing Urdu Large Language Models via Multilingual Synthetic Data Distillation
Muhammad Ali Shafique, Hamza Farooq et al.
We introduce Alif-1.0-8B-Instruct, a multilingual Urdu-English LLM built on Llama-3.1-8B to address low-resource language challenges. Using a modified self-instruct method, we created Urdu-Instruct, a culturally aligned synthetic dataset. Alif outperforms larger multilingual models on Urdu-specific tasks under a $100 training budget.
Muhammad Ali Shafique, Kanwal Mehreen, Muhammad Arham, Maaz Amjad, Sabur Butt, Hamza Farooq
Query Attribute Modeling: Improving Search Relevance with Semantic Search and Metadata Filtering
Karthik Menon, Hamza Farooq et al.
We propose Query Attribute Modeling (QAM), a hybrid framework that transforms free-text queries into structured metadata and semantic components. Tested on the Amazon Toys Reviews dataset, QAM achieved a mAP@5 of 52.99%, surpassing BM25, semantic, and hybrid search baselines.
Karthik Menon, Batool Arhamna Haider, Muhammad Arham, Kanwal Mehreen, Ram Mohan Rao Kadiyala, Hamza Farooq
Robust and Fine-Grained Detection of AI-Generated Texts
Ram Mohan Rao Kadiyala, Hamza Farooq et al.
We present token classification models for detecting human-LLM co-authored texts across 23 languages. Trained on a new dataset of 2.4M samples, our models generalize well to unseen domains, generators, and adversarial inputs.
Ram Mohan Rao Kadiyala, Siddartha Pullakhandam, Kanwal Mehreen, et al., Hamza Farooq
Interested in collaborating?
We work with universities, research labs, and enterprise teams on applied ML research.
Get in touch →