Model Refinement Process
How Refinement Works
Model refinement updates your scoring model based on accumulated feedback. The process:
- Feedback accumulates -- team members submit corrections on lead evaluations
- User triggers refinement from the Training Models page (requires Owner, Admin, or Editor role)
- Preview model is generated -- AI analyzes all pending feedback alongside the current model
- Test evaluation runs -- the preview model scores feedback leads so you can compare old vs. new scores
- User saves and applies -- new model version is created and becomes active
- Leads are re-evaluated -- leads from the last 30 days are queued for re-scoring with the updated model
Refinement is manually triggered, not automatic. Submitting feedback does not start a refinement on its own.
Triggering a Refinement
Prerequisites
- Owner, Admin, or Editor role
- At least 1 pending feedback submission
- Available refinement capacity on your plan
The Refinement Flow
On the Training Models page, a notification appears when pending feedback is available:
- Click Preview Refinement to open the refinement modal
- Review pending feedback -- see each feedback item with the lead, account, feedback text, and submission date
- Click Generate Preview Model -- AI creates a candidate model incorporating your feedback
- Review the preview -- see the refinement summary and model summary describing what changed
- Click Test on Feedback Leads -- run the preview model against feedback leads and compare old scores (Old Status) vs. new scores (New Status)
- Click Save & Apply Model -- creates a new model version and triggers re-evaluation of leads from the last 30 days
What Happens During Re-evaluation
After committing a refinement:
- A new model version is created and set as active
- The previous model version is set to Inactive
- Leads evaluated within the last 30 days are queued for re-scoring
- Score changes (Cold to Warm, Hot to Warm, etc.) appear as leads are re-evaluated
- Score history is preserved
Re-evaluation is queued, not instant. Results typically appear within hours depending on the number of leads.
Refinement Limits
Refinements are metered by your plan. The Training Models page displays:
- Refinements used vs. limit for the current billing cycle
- Remaining refinement credits
When the limit is reached:
- The Preview Refinement button is disabled
- A message indicates all refinement credits have been used and suggests upgrading the plan
- Feedback submission is also blocked on the lead detail page until capacity is restored
Viewing Model History
The Training Models page lists all model versions with:
| Column | Description |
|---|---|
| Status | Active or Inactive |
| Name | Model name (or display name / version fallback) |
| Version | Model version ID |
| Created On | When the model was created |
| Actions | Set as Active (for inactive models), View Details |
Getting the Most from Refinement
Batch feedback before triggering. A single refinement with 10 feedback items produces better results than 10 individual refinements with 1 item each. Wait until you have a meaningful set of corrections.
Focus on clear misscores. Feedback on borderline cases is less useful than feedback on obviously wrong scores. Prioritize leads where the score is clearly misaligned with your criteria.
Review test results before committing. The test evaluation step exists specifically so you can verify the refinement moves scores in the right direction before it goes live.
Next Steps
- Evaluation Feedback -- Submit corrections on lead scores
- Usage Tracking -- Monitor refinement capacity
- ICP Overview -- Revisit your ICP definition