From Synthetic Examples to Production Signals
summary
Claims a usable pipeline that turns synthetic multimodal examples into production training signals using NeMo Data Designer and NeMo Anonymizer, with models from OpenAI/OpenRouter/NVIDIA and on‑prem Neatron. Reports processing ~2.6 trillion tokens across ~1.2–2M jobs, a search accuracy jump to 31.28% from 0% and a 2.2× lift on long‑document understanding after 2M records (45M tokens). Shows reproducible preview/create jobs, LLM judges to filter quality, image-context columns for VLM QA, and anonymizer replacing PII more effectively than regex.
AI-written summary · about
chapters (17)
- 0:00Introduction
- 2:29Synthetic data goals
- 7:02How large language models are trained
- 10:08Why synthetic data
- 11:31NeMo Data Designer overview
- 16:23Data Designer features
- 23:27Validating data pipelines
- 31:25Jupyter Lab setup
- 32:07Text QA pipeline
- 45:27Model providers config
- 57:10Scaling data generation
- 59:17Visual QA pairs
- 1:02:44Images and multimodal
- 1:14:51Smaller models for data
- 1:18:52NeMo Anonymizer intro
- 1:25:13Anonymizing logs
- 1:28:26Q&A
related talks
23mNeMo: Open Source Toolkit for Conversational AI
21mFrom 0 to a crowd sourced neural network in a jiffy
16mProduction MLOps for Spam Detection: Real-Time Multi-Modal AI
30m🤯 No data? No problem! Synthetic data to the rescue
36mNVIDIA NeMo: Toolkit for Conversational AI
17mAI/ML demo derby
29mBuild Customize and Deploy LLMs At-Scale on Azure with NVIDIA NeMo
26mA Practical Introduction to Productionizing NLP Models