NVIDIA Datasets

Open datasets for training coding agents on full trajectories: plans, tool calls, and test runs, plus NVIDIA Nemotron datasets to start from.

Which open fine-tuning datasets make a bank assistant fluent in finance terminology, and how to layer public data with your bank's own.

Open NVIDIA instruction-tuning datasets for single-turn code edits: Nemotron-SFT-SWE-v2, instruction-following chat data, and what to skip.

Open NVIDIA datasets for RL post-training: Nemotron-RL agentic sets, Cascade-RL, and HelpSteer preference data for reward signals.

Open datasets to fine-tune a support assistant that answers from your documents, cites sources, and stays grounded across multi-turn conversations.

Open datasets for multi-attribute reward models: how NVIDIA's HelpSteer family rates helpfulness, correctness, and verbosity separately.

NVIDIA publishes the SFT and RL datasets behind its Nemotron post-training pipeline as open downloads, so the exact training mix can be inspected.

Open datasets to close jailbreak and prompt-injection gaps: NVIDIA Nemotron safety, adversarial, and agentic datasets for fine-tuning and guardrails.

Open datasets with reference answers, like NVIDIA's OpenMathInstruct and Nemotron RL math collections, provide verifiable rewards for RL-based math tutoring training.

Which open SFT datasets to use for Japanese and Portuguese alongside English: multilingual reasoning data, Nemotron personas, and Cascade SFT corpora.