## Line Chart: ARC-C Accuracy vs Training Data Percentage
### Overview
The chart compares the accuracy of four different training approaches (Step-Level Online, Instance-Level Online, Step-Level Offline, and SFT Baseline) across varying percentages of training data (10-50%). Accuracy is measured on the y-axis (55-75), while training data percentage is on the x-axis (10-50%).
### Components/Axes
- **X-axis**: Training Data % (10, 20, 30, 40, 50)
- **Y-axis**: Accuracy (55-75, increments of 5)
- **Legend**: Located at bottom-left, with four entries:
- Green line with stars: Step-Level (Online)
- Blue line with triangles: Instance-Level (Online)
- Yellow line with stars: Step-Level (Offline)
- Purple dashed line: SFT Baseline
- **Title**: "ARC-C" at top-center
### Detailed Analysis
1. **Step-Level (Online)** (Green):
- Starts at 72.2% accuracy at 10% training data
- Peaks at 76.4% at 30% training data
- Slight decline to 75.8% at 50% training data
- Trend: Initial sharp increase followed by stabilization
2. **Instance-Level (Online)** (Blue):
- Begins at 66.5% at 10% training data
- Rises to 73.3% at 30% training data
- Peaks at 75.2% at 40% training data
- Drops to 73.4% at 50% training data
- Trend: Gradual improvement then decline
3. **Step-Level (Offline)** (Yellow):
- Starts at 69.2% at 10% training data
- Peaks at 70.8% at 20% training data
- Declines to 66.5% at 50% training data
- Trend: Early peak followed by steady decline
4. **SFT Baseline** (Purple dashed):
- Constant at 60.6% across all training data percentages
- Trend: Flat line with no variation
### Key Observations
- Step-Level (Online) consistently outperforms all other methods
- Instance-Level (Online) shows the most significant improvement between 10% and 40% training data
- Step-Level (Offline) underperforms compared to its online counterpart
- SFT Baseline remains the lowest-performing method throughout
- All methods show diminishing returns beyond 30-40% training data
### Interpretation
The data demonstrates that online training approaches (both step-level and instance-level) significantly outperform offline methods and the SFT baseline. The Step-Level (Online) method achieves the highest accuracy (76.4%) at 30% training data, suggesting optimal performance at this threshold. The Instance-Level (Online) method shows the most dramatic improvement (from 66.5% to 75.2%) as training data increases, indicating strong scalability. The consistent decline in Step-Level (Offline) performance highlights the importance of online training dynamics. The flat SFT Baseline suggests it represents a fundamental lower bound for this particular task. The diminishing returns observed across all methods beyond 40% training data implies potential inefficiencies in utilizing larger datasets without proportional accuracy gains.
