ML Training

The ML Training dashboard lets BECI administrators monitor and manage SigLIP image similarity model fine-tuning runs for visual inspection classification. You can view training progress, review model metrics, browse training pair data, and deploy approved models.

BECI administrators only. This feature is only visible to users with BECI admin access. If you don't see the ML Training tile on the Home screen, contact your administrator.

Accessing ML Training

1. From the Home screen, click the ML Training tile
2. The dashboard loads with statistics cards at the top and two tabs below: Training Runs and Training Pairs

Statistics Overview

Four summary cards appear at the top of the page:

Card	What It Shows
Total Runs	Total number of training runs across all categories
Deployed Models (green)	Number of models currently deployed to production
Pending Review (yellow)	Models that finished training and need human review before deployment
Training Pairs (blue)	Total image-caption pairs available across all categories

Training Runs Tab

The Training Runs tab shows a table of all model training runs.

Runs Table

Column	Description
Run ID	Unique identifier for the training run (monospace)
Category	The image classification category being trained
Status	Current status badge: queued, training, evaluating, review_needed, deployed, or failed
Metrics	F1 score (if available), or "-" if training is still in progress
Created	When the training run was initiated
Actions	View button to open the run detail modal

Run Detail Modal

Click View on any run (or click the table row) to see full details:

• Metrics — four cards showing Precision, Recall, F1 Score, and Accuracy (each as a percentage, 0-100% scale)
• Training Configuration — Base Model, Epochs, Batch Size, Learning Rate, Training Pairs count, Negative Pairs count
• Training Duration — displayed as hours and minutes
• Error Message — shown only if the run failed

Deploying a Model

When a training run reaches review_needed status, two action buttons appear in the detail modal:

• Deploy Model — approves the model and pushes it to production
• Retrain with More Data — sends the model back for additional training with more pairs

Training Pairs Tab

The Training Pairs tab lets you browse the image-caption pairs used for training, organized by category.

Category List

• Use the search box to filter categories by name
• Use the multi-select checkboxes to filter to specific categories, then click Apply Filter
• Each category card shows: name, a progress bar, total pairs count, and unused pairs count
• Click the arrow icon on any category to view its training pairs

Viewing Training Pairs

Click a category name to open the pairs modal. It displays a grid of cards, each showing:

• The training image (or a placeholder if no image URL is available)
• The caption text below the image

The modal header shows the category name and the total number of pairs loaded.

Troubleshooting

I don't see the ML Training tile

Solutions:

• This feature requires BECI admin access. Contact your administrator if you believe you should have access.

No training pairs loaded yet

Solutions:

• Training pairs need to be loaded via the backend script: python scripts/load_classified_training_pairs.py --environment prod
• Contact the engineering team if you need pairs loaded for a new category.

Training run shows "failed" status

Solutions:

• Click View to open the run detail — the error message will explain what went wrong
• Common causes: insufficient training pairs, memory limits, or infrastructure issues
• Contact the engineering team with the Run ID for further investigation

Related Articles

• Tenant Settings and Branding
• Documents and AI Chat

---

Need help? Contact our support team at [e-mail]