Probe Monitoring
The Probe page provides real-time visibility into which SAE features activate during inference — without modifying the model's output.
How It Works
When monitoring is enabled, miLLM:
- Captures the residual stream activations at the SAE's hooked layer
- Encodes them through the SAE to get feature activations
- Records the top-K most active features per forward pass
- Emits results via WebSocket for real-time display
Controls
| Control | Description |
|---|---|
| Enable/Disable | Toggle monitoring on or off |
| Pause/Resume | Temporarily freeze the display while keeping monitoring active |
| Top-K | Number of top features to track: 5, 10, 20, 50, or 100 |
Live Activations Chart
A bar chart showing the most recently activated features:
- X-axis: Feature index
- Y-axis: Activation magnitude
- Updates in real-time with each inference request
Statistics Panel
Aggregated statistics for each monitored feature:
- Count: Number of times the feature appeared in top-K
- Mean, Min, Max, Std: Activation value statistics
Activation History
A table of recent activation records showing:
- Timestamp of the inference request
- Request ID for correlation
- Top features that activated
- Token position in the sequence
Use Clear History to reset the buffer.
Workflow
Enable monitoring, then send inference requests via the OpenAI API or Open WebUI. Watch which features light up for different prompts — this helps identify which features to steer.