Skip to content

2026

Tuning NeMo Safe Synthesizer for Agent Conversations (Part 2)

In Part 1, we turned 8,034 customer-service conversations into synthetic training data with NeMo Safe Synthesizer, and models trained on it kept 90.4% of the utility of models trained on the real data. That result used the default settings, which raises some natural follow-up questions. Is fine-tuning really necessary, or would a few example conversations in a prompt be enough? Which settings shift the balance between utility and replay of the training data? And can synthetic data help with labels that barely show up?

Part 2 answers those questions with three experiments on the same dataset.

Synthesizing Agent Conversations with NeMo Safe Synthesizer (Part 1)

Every time a customer chats with a support agent or a user asks an AI assistant to get something done, a detailed trace is left behind. It captures what the person wanted, how the agent replied, and which actions it took along the way. As agentic applications take off, these traces are quickly becoming some of the most valuable data a team can learn from. They are also some of the most sensitive, since a single conversation can mix personal details with a business's internal playbook.

That combination makes conversation data a natural candidate for synthetic data, but it raises a practical question. NeMo Safe Synthesizer learns from tables, and a conversation is not a table. Fortunately, it does not need to start as one. With a little reshaping, chat transcripts, event logs, transaction histories, and many other ordered records fit the format NeMo Safe Synthesizer expects.

Speeding Up NeMo Safe Synthesizer Generation with Prompt-Aware Token Budgets

NeMo Safe Synthesizer has expanded to more models and larger context windows. That flexibility helps with longer rows, richer text columns, and time-series groups. However, it also exposed a generation bottleneck. Small jobs and short-row datasets could still pay for decode budgets sized for the full context window.

The updated generation path removes that waste in two places. It starts with a small prompt probe before scaling up the batch size, and it replaces full-context decode caps with prompt-aware caps empirically derived from token lengths of the actual data.

Our experiments show up to 10.0x generation speedup, a 1.62x median speedup across 39 before/after comparisons, and more consistent generation times overall.

This improvement shipped in NeMo Safe Synthesizer v0.0.6.

Does NeMo Safe Synthesizer Actually Work? A Case Study with Financial Transactions Data

NeMo Safe Synthesizer creates private, safe versions of sensitive tabular datasets: entirely synthetic data with no one-to-one mapping to the original records, while preserving the statistical structure to remain useful for downstream AI and analytics.

That promise sounds simple, but it raises the question every synthetic data system eventually has to answer: does it actually work?

Private by Design: Introducing NeMo Safe Synthesizer

Every organization working on AI faces the same challenge: the data that would make their models most useful is also proprietary data with the highest barriers to access. The data is right there: patient records, financial transactions, customer support logs, and datasets full of names, account numbers, and personal details. It is rich and perfectly suited to the task, but legal and compliance teams have marked it off-limits for good reason.

We built NeMo Safe Synthesizer to break that deadlock by helping organizations create synthetic versions of sensitive tabular data.