Data for Agents

Source: HuggingFace Date: 2026-07-08 URL: https://huggingface.co/blog/nvidia/open-data-for-agents

Summary

NVIDIA publishes a large batch of synthetic Nemotron datasets aimed specifically at agent training rather than general pretraining: Nemotron-CC-v2 (8.79B samples), Nemotron-CC-Math-v1 (190M samples), 15 pretraining collections totaling over 10 trillion tokens, and Nemotron-Personas (2.4B+ synthetic personas across 10+ countries and languages). The stated target is the hard part of agent capability — software-engineering traces, tool-use failures, multi-step reasoning, retrieval, safety, and user simulation — packaged as synthetic data so organizations can share useful training signal “without exposing the underlying sources.” All datasets are on Hugging Face.

Implications

Feeds the open-weight wave thread from the infrastructure layer rather than the model layer: this is NVIDIA seeding the raw material that lets smaller or newer labs train agent-capable open-weight models without each one having to independently solve data collection for tool-use and multi-step reasoning traces. Combined with the SWE-bench integrity problems surfacing the same day (see the OpenAI eval signal), it’s a reminder that both training data and evaluation for “agentic coding capability” are unusually unsettled right now — the wave of open weights claiming agent competence is being built on infrastructure (data, benchmarks) that’s still being actively contested and rebuilt underneath it.