英偉達公開10萬億訓練資料. 本文保留已核驗事實，並以產品、基礎設施與治理脈絡重新整理。
發生了什麼
英偉達公開10萬億訓練資料 is the current news peg. The verified record centers on these points:
- NVIDIA and Hugging Face presented Nemotron open data for agents
- the headline corpus exceeds 10 trillion pre-training tokens plus millions of post-training samples
- Prompt Atlas visualizes sampled prompts by dataset, pipeline stage, domain and tool use
- one data card lists 6,585.8B tokens with English Common Crawl, synthetic data, QA, math and code components
為何重要
更大的產業訊號是 agent training needs auditable data shapes for tool calls, failure recovery and workflows, not only open weights.
接下來看什麼
接下來要看的重點是 how teams document synthetic data generation, filtering, licensing and workflow coverage before enterprise deployment.
參考來源：Hugging Face/NVIDIA Data for Agents, NVIDIA Blog, Hugging Face Nemotron data cards and collections, CocoLoop.