NeMo Data Designer

NVIDIA's open-source framework for synthetic data generation, used to build targeted training data for the Nemotron models. A dataset is declared column by column: samplers that steer diversity (categories and statistical distributions), LLM-generated text, code and structured outputs, images, embeddings, seed data, and validators or LLM judges that score each record. Data Designer (NDD) resolves the dependencies between columns, schedules asynchronous calls to user-supplied model endpoints (hosted APIs, gateways or OpenAI-compatible servers), retries failed requests, and can call MCP tools while keeping their traces. The configuration is itself an inspectable, shareable artifact, and the core loop is to preview a few records, revise the specification, then run it at full scale; plugins add column types, seed readers and processors. The repository was created in October 2025 and tagged v0.1.0 in November 2025; this entry is dated to the technical report posted to arXiv on September 15, 2026, which describes version 0.9.0. Apache 2.0, installed with pip install data-designer, about 2,300 GitHub stars.

The report's case studies document how NVIDIA has used it. About 9,000 JSON-schema tasks built with NDD went into the RL stage of Nemotron 3 Nano (the released set holds 9,949 verified examples) and lifted JSONSchemaBench from 80.2% to 86.9% and StructEval-Text from 64.5% to 72.1%; a usability curriculum for Nemotron 3 Ultra, developed with Perplexity, raised StructEval-T from 78.6% to 82.1%. A text-to-SQL pipeline generated 300K candidate queries across PostgreSQL, MySQL and SQLite and kept 96.5K, and the report cites a BIRD gain from 26.77% to 41.80% against 38.25% for GPT-OSS-120B. A search-agent pipeline turned 50K Wikidata paths into 24K obfuscated multi-hop questions and about 7K validated trajectories averaging 12 tool calls, which went into Nemotron 3 Super SFT. A long-document vision pipeline produced about 11.4M visual question-answer pairs (about 45B tokens), with development checkpoints rising from 26.32% to 59.00% on MMLongBench-Doc. Prompt-variation data roughly halved Nemotron 3 Nano's sensitivity to prompt wording across GPQA, MMLU-Pro, competition math and LiveCodeBench, and the census-grounded Nemotron-Personas sets span 10 regional datasets in 15 language or script editions, covering regions home to about 2.4B people. Outside NVIDIA, CrowdStrike used NDD to write descriptions for analyst queries and fine-tuned Llama-3.3-Nemotron-Super-49B-v1.5 to 96% valid-query accuracy, against 94% for Claude Sonnet 4.5.