面向 Nemotron 预训练的任务种子合成 QA 数据生成

This article is a deep-dive from JudyAI Lab — an AI engineering playbook series with 100+ published guides, 5,000+ weekly readers across 60+ countries, focused on the practical side of running AI agents, trading systems, and content pipelines in production. 📰 Key Takeaways NVIDIA built a five-stage "Task-Seeded SDG" (Synthetic Data Generation) pipeline for the Nemotron model family: it pulls roughly 70 public tasks (about 700 subtasks) from lm-eval-harness, split into knowledge-intensive (39 tasks, ~3M samples) and reasoning-intensive (34 tasks, ~1.5M samples) seed categories, uses a large language model to generate QA pairs that differ in content but match the seed tasks' skill level, then appends reasoning chains and domain knowledge before uniformly filtering and packaging the data. In ablation experiments, the context-enriched version won decisively: GPQA-Diamond CoT rose from 34.85 to 45.96 (+11.11), AGIEval-en CoT gained +6.16, and MMLU-Pro 5-shot gained +2.44. Mixing this synthetic data into Nemotron-3 Nano's post-training (at the ~100B token scale) ultimately pushed GPQA from 30.8 to 41.9 (+11.1), with MMLU-Pro +1.8, coding ability +1.9, and common sense understanding +1.6 — multiple dimensions improving together, confirming that broad task coverage effectively prevents overfitting to a single evaluation style. Key design principles include: answers should be stored as semantic text rather than option letters, and when mixing datasets, task ratios must be carefully balanced to ensure stable, across-the-board gains in knowledge, reasoning, and coding capability. 💬 JudyAI Lab Take NVIDIA's five-stage "Task-Seeded SDG" pipeline for the Nemotron family is the first concrete demonstration of how a structured method can produce training data at scale — letting a small model improve across multiple evaluations at once, instead of just chasing a high score on one task. What's most worth paying attention to here is the deliberate split between "knowledge-intensive" and "reasoning-intensive" seed tasks, and the careful balancing of task ratios when mixing them into post-training. The ablation results are clear: the context-enriched version pushed GPQA-Diamond CoT from 34.85 to 45.96, a gap of over 11 points. That tells us synthetic data quality isn't just about generation volume — it's about structural design. Broad coverage across roughly 70 public tasks and 700 subtasks is exactly what prevents the model from overfitting to a specific evaluation style. The fact that coding ability, common sense understanding, and reasoning ability all improved together shows that task coverage breadth is itself an overfitting defense. One other detail worth remembering: answers should be stored as semantic text rather than option letters, so the model actually learns semantic understanding instead of memorizing option position. If you're adding synthetic training data to your own model or application, it's worth asking first: are my task seeds diverse enough, or am I only betting on a single capability dimension? 📅 Source Info Published: 2026-06-04T11:24 Original source: https://huggingface.co/blog/nvidia/task-seeded-sdg 🔗 Further Reading The Rise of Personalized AI Models: Tailoring Intelligence for Your Business From Trading Idea to Live Execution: The Real Workflow of AI-Assisted Strategy Development References Nemotron AI Models | NVIDIA Developer NVIDIA Nemotron: Advanced Multimodal AI Models for Agentic Reasoning Nemotron 3 Nano - A New Standard for Efficient, Open, and Intelligent Agentic Models - Hugging Face Docs Originally published at Judy AI Lab. Visit for more articles on AI engineering and development.