Fine-tuning

A guided pipeline to take your project’s classification history and produce a fine-tuned variant that beats the current production prompt.

Pipeline

exporting → exported → training → trained → validating → validated → promoted
                                                                  ↘
                                                              rejected

Each stage is observable in the UI with a stepper. The Validation report includes:

• Eval accuracy vs. baseline.
• Severity calibration delta.
• Component tagging F1.
• Cost-per-classification projection.

A candidate is only promotable if its judge-score mean wins the active prompt’s mean by ≥ 0.05 with p < 0.05. Otherwise the UI offers Reject with a one-line reason, archived for future tuning runs.