Synthetic SFT data

problem

a tool-calling assistant needs training dialogues that match what its backend really returns, and hand-written ones drift from that.

what I built

• a pipeline that writes multi-turn dialogues grounded in real backend tool outputs
• multi-tool routing driven by versioned YAML constraint prompts
• regex validators with corrective retry loops, in place of hard-coded output patches

results

• synthetic artifacts get caught and retried instead of leaking into the training data