Open datasets for a support assistant that cites sources across turns

Build the skill mix one behavior at a time

Use three NVIDIA RL datasets for the core behaviors: Nemotron-RL-Instruction-Following-Citation-Formatting-v1 for citing sources in a required format, Nemotron-RL-Instruction-Following-MultiTurnChat-v1 for following instructions across turns, and Nemotron-RL-QA-Abstention-v1 for declining to answer when the evidence is missing. For supervised fine-tuning, the knowledge_agent subset of Nemotron-SFT-Agentic-v4 contains conversations in which an assistant uses tools and policy or task documents to respond to users. A support assistant has to combine these skills in every reply. It must answer only from retrieved documents and attach the right source to each claim, while it tracks what the user said earlier. One dataset rarely teaches all of that, so build the mix behavior by behavior. The tool_calling subset of Nemotron-SFT-Agentic-v4 also helps when the assistant must look up order or account details through tools.

Public data teaches general grounding behavior. Your own documents teach the domain. Generate synthetic multi-turn dialogues over your knowledge base with NVIDIA NeMo Data Designer, require a citation in every assistant turn, and include questions your documents cannot answer. Keep the retrieved passages in each training example, so the model learns to cite the text it actually received. Nemotron-SFT-Instruction-Following-Chat-v4 adds chat and multi-turn instruction-following data, including response-length control, which keeps the assistant useful outside the core support flow. Before launch, hold out real support questions and measure citation accuracy and correct abstention as separate scores.

Datasets to start with

• Nemotron-RL-Instruction-Following-Citation-Formatting-v1: RL data for citing sources in a required format.

• Nemotron-RL-Instruction-Following-MultiTurnChat-v1: RL data for following instructions across conversation turns.

• Nemotron-RL-QA-Abstention-v1: RL data for knowing when not to answer.

• Nemotron-SFT-Agentic-v4: tool-use SFT data whose knowledge_agent subset answers from policy and task documents.

Key takeaways

• Grounded answers, citations, multi-turn context, and abstention are separate skills. Train each one.

• NVIDIA publishes RL data for citation formatting, multi-turn instruction following, and QA abstention.

• The knowledge_agent subset of Nemotron-SFT-Agentic-v4 shows assistants answering from policy and task documents.

• Synthetic dialogues over your own documents close the domain gap.

• Score citation accuracy and abstention separately on held-out real questions.