Generating Training Data

Generate synthetic task data (user queries) for the Workplace Assistant environment using NeMo Data Designer.

This pipeline focuses on generating tasks for use with the environment. It also simulates agent trajectories, but these are used for quality filtering and validation — the environment itself produces the actual model responses during rollout collection. The Workplace Assistant uses 27 tools across 6 databases, and NeMo Data Designer can produce realistic multi-step user queries at scale.

Pipeline Overview

The data generation pipeline:

1. Load tool schemas for the Workplace Assistant environment
2. Use NeMo Data Designer to generate realistic multi-step user queries
3. Simulate agent trajectories (step-by-step tool-call solutions)
4. Apply dual-level LLM judge filtering to ensure data quality
5. Export task data in NeMo Gym JSONL format

Notebook

The tutorial is provided as a Jupyter notebook. See the notebook README for prerequisites and setup instructions.

View Notebook on GitHub

What’s Next?

After generating your task data, use it with the Workplace Assistant resources server to collect rollouts (where the environment produces model responses) and then proceed to GRPO training.