## Box Plots: Depthwise Average MIN-K% ### Overview The image contains four box plots (a–d) comparing "Depthwise Average MIN-K%" across three depths (Depth 1, 2, 3) for different LLaMA models and tasks. Each plot uses distinct colors (blue, orange, green) for visual differentiation. ### Components/Axes - **X-axis**: Depth (1, 2, 3) - **Y-axis**: Depthwise Average MIN-K% (range: 0–6) - **Legends**: No explicit legend for box plots; colors (blue, orange, green) are used for visual distinction. - **Titles**: - (a) LLaMA 2 7B Chat - (b) LLaMA 2 70B Chat - (c) LLaMA 3 8B Instruct - (d) LLaMA 3 70B Instruct ### Detailed Analysis - **Box Plot Trends**: - **Depth 1**: Medians range from ~3.5 (a) to ~4.5 (d). - **Depth 2**: Medians increase slightly (e.g., ~4.0 in a, ~4.5 in d). - **Depth 3**: Medians rise further (e.g., ~5.0 in a, ~5.5 in d). - **Outliers**: Scattered points above the upper whiskers, indicating variability. - **Color Consistency**: Colors (blue, orange, green) are consistent across plots but lack a legend. ### Key Observations - **Depth Correlation**: Higher depths (3) generally show higher MIN-K% values, suggesting improved performance with depth. - **Model/Task Variability**: Larger models (e.g., 70B) and instruction-tuned variants (e.g., 3 70B Instruct) exhibit higher medians. ## Density Plots: Score Gap (D3-D2) ### Overview Three density plots (e–g) visualize the distribution of score gaps between Depth 3 (D3) and Depth 2 (D2) for the same models/tasks as the box plots. ### Components/Axes - **X-axis**: Score Gap (D3-D2) (range: -1.5 to 1.5) - **Y-axis**: Density (0–1.25) - **Legends**: - Green: 25% percentile - Orange: 75% percentile - **Titles**: - (e) LLaMA 2 7B Chat - (f) LLaMA 2 70B Chat - (g) LLaMA 3 70B Instruct ### Detailed Analysis - **Density Trends**: - **25% (Green)**: Peaks near 0.5–1.0, indicating most score gaps are positive. - **75% (Orange)**: Peaks at higher values (e.g., ~1.0–1.5), showing a skew toward larger gaps. - **Distribution**: Narrower peaks in (e) and (f) vs. sharper, taller peaks in (g), suggesting more concentrated gaps in larger models. ### Key Observations - **Positive Gaps**: Most score gaps are positive, implying D3 outperforms D2. - **Model-Specific Patterns**: Larger models (e.g., 70B) show sharper distributions, indicating more consistent performance improvements. ## Interpretation The data demonstrates that deeper layers (Depth 3) consistently yield higher MIN-K% scores across models and tasks, suggesting improved performance with depth. The score gaps (D3-D2) are predominantly positive, with larger models (e.g., 70B) exhibiting more pronounced and concentrated improvements. The 75% percentile gaps are significantly higher than the 25% percentile, highlighting that a majority of the data lies in the upper range of performance gains. This aligns with the box plot trends, where Depth 3 consistently outperforms shallower depths. The absence of a legend for the box plots limits direct interpretation of color coding, but the density plots’ legend clarifies the percentile distributions.