← Back to Dashboard
1. When to Customize a Model2. Evaluation of Customized Models

Evaluation of Customized Models

📚 Fine-Tuning and Customization11 min90 XP

Quality and Safety Gatekeeping

Evaluate customized models on task quality, safety behavior, latency, and cost impact. A quality gain that increases risk is not a production win - fine-tuning can shift safety behavior in ways base-model evaluations never covered.

The Four-Way Comparison

A customized model must beat all realistic alternatives, not just its own base model:

AlternativeQuestion it answers
Base model + your best promptWas prompting alone enough?
Base model + RAGWas missing knowledge the real gap?
A stronger off-the-shelf tierWould upgrading the model beat customizing this one?
Your fine-tuneDoes customization win after Provisioned Throughput cost?

Evaluation Protocol

  • Held-out test sets - never evaluate on training data; keep a frozen split the training pipeline cannot see.
  • Adversarial prompts - fine-tuning can erode refusal behavior; rerun the full safety suite, not just task metrics.
  • Regression breadth - test adjacent tasks too: a model tuned for triage summaries may get worse at general Q&A your app also uses.
  • Latency + cost - measure full with Provisioned Throughput economics included, not just per-token price.

Post-Deployment Drift

Track drift after deployment: input distributions shift, and a model tuned on last quarter's ticket styles slowly degrades. Schedule periodic re-evaluation on fresh samples and define a retirement/retraining trigger (e.g. task success −5% versus launch baseline).

Decision rule: promote a custom model only when it wins the four-way comparison on quality and shows zero safety regressions and the total cost of ownership (training + Provisioned Throughput + re-evaluation cadence) beats the best alternative. Two out of three is a rejection.
🧪 Knowledge Check
Press 1-4 to select1 of 2
A valid model customization launch requires:
Higher quality only
Quality + safety + latency + cost validation
No baseline comparison
No monitoring