NVIDIA 发布 NeMo Data Designer 合成数据工具报告
开源合成数据工具用声明式配置定义数据列，支持预览后全量生成，Nemotron 案例中 JSON 任务准确率从 80.2% 升至 86.9%。
Banger release from NVIDIA. They just published a report on NeMo Data Designer, their open-source synthetic data tool. It's a declarative config which makes a synthetic data pipeline easy to review, share and rerun. In NDD, a person or an agent defines each dataset column in a config file. Column types include generated text, code, structured output, images, embeddings and statistical samplers that control diversity. Plugins add new types. The workflow is built around previews. You generate a few records, check them, adjust the config and then run at full scale. The runtime handles column dependencies, calls to your model endpoints and retries. The paper includes Nemotron case studies. About 9K JSON-schema tasks made with NDD raised Nemotron Nano v3 from 80.2% to 86.9% on JSONSchemaBench. Paper: https://t.co/v43t8JEWTI