Skip to main content

👑 Admin Guide

Administrative operations for managing datasets, AI training cycles, and system monitoring.

Admin Guide​

This guide explains how administrators manage:

  • Dataset review
  • Periodic training (Option Periodic)
  • Model registry
  • System monitoring

Admin access requires role: Admin.

1. Admin Pages Overview​

1️⃣ Admin Dataset​

Purpose: Review collected URLs before they are used for retraining.

Features:

  • View dataset entries by status (Pending / Approved / Rejected)
  • Approve URL samples
  • Reject incorrect/noisy samples
  • Export dataset to CSV
  • Filter/search URLs

2️⃣ Admin Train Jobs​

Purpose: Manage periodic training process.

Features:

  • Trigger training manually
  • View job history
  • Monitor job status (Running / Completed / Failed)
  • View evaluation metrics
  • Confirm model activation

2. Dataset Workflow​

1

Step 1 — Automatic Collection​

When a URL is scanned:

  • System logs it in ScanLogs
  • System inserts entry into UrlDataset
  • Default status: Pending
2

Step 2 — Admin Review​

Admin reviews each Pending record.

Possible actions:

Approve

  • URL considered valid
  • Included in next training cycle

Reject

  • URL excluded from training
  • Marked as noisy / incorrect / duplicate

Only Approved entries affect model retraining.

3

Step 3 — Export​

Admin can export dataset to CSV for:

  • Offline inspection
  • External ML experiments
  • Audit compliance
  • Research documentation

3. Training Workflow (Option Periodic)​

Periodic training ensures model stability and safety.

Trigger Methods​

  • Manual trigger from Admin UI
  • Scheduled background job (daily / weekly)

Training Process​

  1. Load baseline dataset

  2. Merge approved dataset entries

  3. Perform feature extraction

  4. Train hybrid model:

    • Random Forest (URL features)
    • NLP pipeline (TF-IDF)

  5. Evaluate metrics:

    • Accuracy
    • Precision
    • Recall
    • F1 Score

  6. Save model artifact

  7. Register model in ModelRegistry

4. Evaluation & Quality Gates​

Before activation, model must satisfy:

  • High precision for block categories
  • Acceptable recall for harmful classes
  • No major regression vs previous version
  • Acceptable false-positive rate

If quality gates fail:

  • Model is not activated
  • Previous version remains active

5. Model Registry​

ModelRegistry stores:

  • Model version
  • Training date
  • Dataset size
  • Evaluation metrics
  • Activation status
  • Notes

Only one model should be active at a time.

6. Monitoring & Audit​

Admins should regularly:

  • Review dataset inflow
  • Monitor false positives
  • Check drift patterns
  • Validate training metrics
  • Verify model activation logs

7. Best Practices​

  • Do not approve noisy samples
  • Avoid training with too few samples
  • Keep baseline dataset stable
  • Document each model version
  • Maintain rollback readiness

8. Security Considerations​

  • Admin endpoints must be protected by role-based authorization
  • All admin actions should be logged
  • Training trigger must not be publicly accessible
  • Dataset manipulation must be audited

1
Review pending dataset

2
Approve clean samples

3
Trigger training

4
Review metrics

5
Activate new model (if qualified)

Summary​

Admin responsibilities include:

  • Dataset governance
  • Safe model lifecycle management
  • Audit tracking
  • Ensuring system stability

The goal is controlled AI improvement, not rapid uncontrolled updates.