Dhruv Nathawani

Dhruv Nathawani works on synthetic training data that teaches language models specific behaviors: following unfamiliar schemas, querying messy databases, and searching for answers they cannot retrieve from memory. His contributions to NeMo Data Designer connect those tasks through a practical approach: design the variation in the examples, generate candidate records, and check which ones deserve to become training data.

From clinical prediction to model training

Dhruv’s earlier research included a collaboration on multimodal learning for prostate-cancer prognosis, whose publication lists his affiliation with Salesforce Research. The team combined pathology images with clinical information from randomized trials to predict patient outcomes. Dhruv was a co-author in a collaboration spanning machine learning and clinical research.

At Gretel, his work included datasets intended to change language-model behavior. With Lipika Ramaswamy and Maarten Van Segbroeck, he co-authored the release of an open synthetic safety dataset containing 8,361 examples. Each record paired a prompt and response with a safer alternative, giving developers concrete examples of the response being corrected and its intended replacement.

His subsequent NVIDIA work includes contributions to Data Designer and synthetic-data pipelines for Nemotron models. These pipelines make different demands on the training examples: a search agent needs a useful sequence of tool calls, a structured-output model must obey a schema it has not seen before, and a SQL model must interpret both the question and the database’s irregularities.

Designing examples around the capability

• Search-agent training: Dhruv’s search-data pipeline starts with paths through connected facts in Wikidata and turns them into indirect questions requiring investigation. Models attempt those questions using live web search through Tavily. The resulting records include search calls, intermediate reasoning, and answers, allowing supervised fine-tuning to teach a method of investigation alongside the final result. Question construction, agent execution, and preparation of the surviving training trajectories remain separate stages.

• Following unfamiliar schemas: In his structured-output work, each training record receives a newly generated schema. Reusing a few fixed structures risks teaching a model to memorize them; varying the schemas targets its ability to follow unfamiliar instructions. The pipeline combines generation with programmatic checks and rejection sampling. NVIDIA reported that training Nemotron Nano v3 with this data increased valid outputs on JSONSchemaBench from 80.2% to 86.9%.

• SQL for messy databases: With Yev Meyer and Maarten Van Segbroeck, Dhruv developed a text-to-SQL dataset pipeline that deliberately introduces dates stored as text, currency symbols embedded in values, and similar-looking tables irrelevant to the question. These complications require the model to select the right schema elements and handle awkward representations as well as produce valid SQL. Dialect-specific validators and multiple model judges filter the generated examples.

Dhruv’s workshop on teaching agents to search makes the cost of that filtering concrete: 50,000 seed examples become 7,000 training records after generation and filtering. He emphasizes previewing small samples before running a batch and distinguishes the evidence supplied by deterministic checks, model judgments, and human review. Even a question grounded in Wikidata can have a stale answer, so a source graph or a favorable model score cannot by itself establish correctness. The process requires checking both what the examples ask and what the generated trajectories teach.

Synthetic data as an experimental tool

Dhruv also co-authored TinySQL, which uses progressively harder synthetic SQL tasks and small models to study how query-generation capabilities develop across model components and layers. Here, constructing the data provides a controlled setting for investigating model internals, extending its role beyond supplying more training examples.

In research on transferring activation-space interventions between models, Dhruv and his collaborators learned mappings between internal representations to transfer interventions for harmful-prompt refusal and backdoor removal. This work examines another way to change model behavior: transferring an intervention through relationships between models’ activation spaces.