toolgenTRAINING DATA ENGINE
SCHEMA-GROUNDED SYNTHETIC DATA

Turn tool definitions into conversations agents can learn from.

Toolgen samples connected endpoints, generates a realistic multi-turn task, executes schema-valid mock tools, and judges the result before it reaches your dataset.

16 tools 40 endpoints 8 categories
Tool schemas connected into a structured conversation trace
Interactive sample · deterministic mock execution
CONVERSATION TRACE

Book a Paris hotel within budget

READY
8 turns · 2 tool calls · disambiguation
QUALITY GATE 4.7/ 5 mean score
Naturalness4.6
Tool correctness4.8
Task completion4.7
UNDER THE HOOD

A small pipeline with strict contracts.

01Build

Normalize ToolBench schemas into a typed endpoint registry and NetworkX graph.

02Sample

Choose a connected tool chain while steering toward underrepresented categories.

03Generate

Planner, user, and assistant agents create a coherent task around that chain.

04Verify

Mock execution grounds IDs and values; the judge scores and repairs weak traces.

OUTPUT

Ready for your JSONL dataset.

Every record includes messages, tool calls, judge scores, seed, category coverage, and repair metadata.