## Line Graph: Log Probability of Output vs. SecAlign Training Steps
### Overview
The graph depicts the evolution of log probability of output during SecAlign training steps, comparing two models (SecAlign and StruQ) across desirable and undesirable output categories. The y-axis uses a logarithmic scale to emphasize differences in probability magnitudes.
### Components/Axes
- **X-axis**: "SecAlign training step(s)" (0–200 steps)
- **Y-axis**: "Log probability of output" (-300 to -50)
- **Legend** (right-aligned):
- Dashed orange: SecAlign (desirable output)
- Solid orange: SecAlign (undesirable output)
- Dashed blue: StruQ (desirable output)
- Solid blue: StruQ (undesirable output)
### Detailed Analysis
1. **SecAlign (undesirable output)**
- Solid orange line starts at ~-100 log probability at step 0.
- Sharp decline to ~-300 by step 50, followed by erratic oscillations between -250 and -300.
- Final value at step 200: ~-280.
2. **SecAlign (desirable output)**
- Dashed orange line remains stable near -50 throughout training.
3. **StruQ (undesirable output)**
- Solid blue line stays constant at ~-150.
4. **StruQ (desirable output)**
- Dashed blue line remains flat at ~-50.
### Key Observations
- SecAlign's undesirable output probability drops dramatically in early training but stabilizes with high variability.
- StruQ's outputs show no change during training, suggesting static performance.
- Desirable outputs for both models remain constant, indicating no improvement in target metrics.
### Interpretation
The data suggests SecAlign prioritizes reducing undesirable outputs early in training, though its desirable output remains unchanged. StruQ's static performance implies it lacks adaptive training dynamics. The logarithmic scale highlights SecAlign's significant improvement in undesirable output suppression, but the lack of progress in desirable outputs raises questions about training efficacy. The oscillations in SecAlign's undesirable output post-step 50 may indicate instability or overfitting risks.
