LangGraph CLI Synthetic Dataset Generation

Why it matters

Synthetic data is only as good as its coverage and its checks. This notebook builds both into the generator, then reminds you that a valid JSON label can still be wrong.

What to practise

Define a CLIToolCall Pydantic schema and build a NeMo Data Designer configuration: category and uniform sampler columns that spread the commands, templates and ports, an LLM text column for the user request, and an LLM structured column whose output_format is the schema. Run designer.preview() on five records before designer.create(), then watch how many rows survive the grounding, empty or length, and duplicate checks. Split the rest 90/10 into train.jsonl and val.jsonl.

How it fits

The samplers, schema and post-generation filters are 6.4.3’s admission of synthetic examples; compare the notebook’s checks with that Bite’s structural, grounding and coverage checks. The duplicate filter is 6.4.1’s deduplication applied to generated text.