# Technical Data Extraction: Mean Accuracy Performance by Compression Method This document provides a detailed extraction of the data presented in the bar chart comparing various model compression methods across four task categories. ## 1. Metadata and Axis Information * **Y-Axis Title:** Mean Accuracy (%) * **Y-Axis Scale:** 0 to 80+ (increments of 20 marked: 0, 20, 40, 60, 80) * **X-Axis Categories (Tasks):** 1. Coding 2. Math 3. Multiple Choice 4. Creative Writing * **Baseline:** A horizontal dashed grey line represents the **0% Compression Ratio** (uncompressed performance) for each task category. ## 2. Legend and Classification The chart uses a combination of color (Method) and texture/shading (Compression Ratio) to display data. ### Compression Ratio (Texture/Shading) * **0%:** Horizontal dashed grey line (Baseline). * **50%:** Bars with diagonal hatching (forward slash pattern). * **25%:** Solid colored bars (placed behind/above the 50% bars). ### Methods (Colors) Each method is evaluated in two contexts: **(specific)** and **(general)**. * **REAP (ours):** Blue (Specific: Medium Blue / General: Dark Grey-Blue) * **EAN:** Red/Pink (Specific: Bright Pink / General: Dusty Rose) * **Frequency:** Green (Specific: Kelly Green / General: Olive Green) * **HC-SMoE:** Yellow/Tan (Specific: Gold / General: Beige) * **M-SMoE:** Light Blue (Specific: Sky Blue / General: Pale Blue) --- ## 3. Data Extraction by Task Category Values are estimated based on the Y-axis scale. Within each method group, the left bar is "(specific)" and the right bar is "(general)". | Task Category | Baseline (0%) | Method | Specific (%) | General (%) | | :--- | :--- | :--- | :--- | :--- | | **Coding** | ~58% | REAP (ours) | ~56% | ~53% | | | | EAN | ~55% | ~32% | | | | Frequency | ~57% | ~47% | | | | HC-SMoE | ~53% | ~54% | | | | M-SMoE | ~56% | ~52% | | **Math** | ~89% | REAP (ours) | ~88% | ~89% | | | | EAN | ~87% | ~86% | | | | Frequency | ~88% | ~89% | | | | HC-SMoE | ~85% | ~86% | | | | M-SMoE | ~89% | ~89% | | **Multiple Choice** | ~72% | REAP (ours) | ~69% | ~70% | | | | EAN | ~65% | ~68% | | | | Frequency | ~65% | ~68% | | | | HC-SMoE | ~67% | ~64% | | | | M-SMoE | ~56% | ~67% | | **Creative Writing** | ~81% | REAP (ours) | ~80% | ~79% | | | | EAN | ~81% | ~79% | | | | Frequency | ~81% | ~78% | | | | HC-SMoE | ~50% | ~58% | | | | M-SMoE | ~81% | ~74% | --- ## 4. Key Observations and Trends * **REAP (ours):** Consistently maintains performance closest to the 0% baseline across all tasks and both specific/general contexts, showing high robustness to 50% compression (hatched areas). * **Performance Degradation:** * **Frequency (general)** and **HC-SMoE (specific)** show catastrophic failure in the "Creative Writing" task at 50% compression (dropping to ~15% and ~4% respectively). * **EAN (general)** shows a significant performance drop in the "Coding" task at 50% compression. * **Task Difficulty:** The "Math" category shows the highest overall accuracy and the least variance between compression methods, while "Coding" and "Creative Writing" show the highest sensitivity to different compression algorithms.

## Bar Chart: Mean Accuracy (%) ### Overview The bar chart compares the mean accuracy of different methods across four domains: Coding, Math, Multiple Choice, and Creative Writing. The methods compared are REAP (ours) (specific), REAP (ours) (general), EAN (specific), EAN (general), Frequency (specific), Frequency (general), HC-SmoE (specific), HC-SmoE (general), M-SmoE (specific), and M-SmoE (general). ### Components/Axes - **X-axis**: Domain (Coding, Math, Multiple Choice, Creative Writing) - **Y-axis**: Mean Accuracy (%) - **Legend**: Compression Ratio (0%, 25%, 50%), Method (REAP (ours) (specific), REAP (ours) (general), EAN (specific), EAN (general), Frequency (specific), Frequency (general), HC-SmoE (specific), HC-SmoE (general), M-SmoE (specific), M-SmoE (general)) ### Detailed Analysis or ### Content Details - **Coding**: The highest mean accuracy is achieved by REAP (ours) (specific) at 60%, followed by REAP (ours) (general) at 55%, EAN (specific) at 50%, EAN (general) at 45%, Frequency (specific) at 40%, Frequency (general) at 35%, HC-SmoE (specific) at 30%, HC-SmoE (general) at 25%, M-SmoE (specific) at 20%, and M-SmoE (general) at 15%. - **Math**: The highest mean accuracy is achieved by REAP (ours) (specific) at 85%, followed by REAP (ours) (general) at 80%, EAN (specific) at 75%, EAN (general) at 70%, Frequency (specific) at 65%, Frequency (general) at 60%, HC-SmoE (specific) at 55%, HC-SmoE (general) at 50%, M-SmoE (specific) at 45%, and M-SmoE (general) at 40%. - **Multiple Choice**: The highest mean accuracy is achieved by REAP (ours) (specific) at 70%, followed by REAP (ours) (general) at 65%, EAN (specific) at 60%, EAN (general) at 55%, Frequency (specific) at 50%, Frequency (general) at 45%, HC-SmoE (specific) at 40%, HC-SmoE (general) at 35%, M-SmoE (specific) at 30%, and M-SmoE (general) at 25%. - **Creative Writing**: The highest mean accuracy is achieved by REAP (ours) (specific) at 75%, followed by REAP (ours) (general) at 70%, EAN (specific) at 65%, EAN (general) at 60%, Frequency (specific) at 55%, Frequency (general) at 50%, HC-SmoE (specific) at 45%, HC-SmoE (general) at 40%, M-SmoE (specific) at 35%, and M-SmoE (general) at 30%. ### Key Observations - REAP (ours) (specific) consistently outperforms other methods across all domains. - REAP (ours) (general) performs well in most domains but slightly below REAP (ours) (specific). - EAN (specific) and EAN (general) show significant performance differences, with EAN (specific) generally outperforming EAN (general). - Frequency (specific) and Frequency (general) have the lowest mean accuracy across all domains. - HC-SmoE (specific) and HC-SmoE (general) show a consistent performance trend, with HC-SmoE (specific) generally outperforming HC-SmoE (general). - M-SmoE (specific) and M-SmoE (general) have the lowest mean accuracy across all domains. ### Interpretation The data suggests that REAP (ours) (specific) is the most effective method across all domains, followed by REAP (ours) (general). EAN (specific) and EAN (general) show significant performance differences, with EAN (specific) generally outperforming EAN (general). Frequency (specific) and Frequency (general) have the lowest mean accuracy across all domains. HC-SmoE (specific) and HC-SmoE (general) show a consistent performance trend, with HC-SmoE (specific) generally outperforming HC-SmoE (general). M-SmoE (specific) and M-SmoE (general) have the lowest mean accuracy across all domains. The compression ratio does not seem to have a significant impact on the mean accuracy of the methods.

## Technical Document Image Analysis ### Image Description The image is a grouped bar chart illustrating the "Mean Accuracy (%)" of various methods across four different tasks: "Coding", "Math", "Multiple Choice", and "Creative Writing". The chart compares the performance of these methods at two different compression ratios (50% and 25%) against a 0% compression baseline. ### Chart Details * **Y-axis:** Labeled "Mean Accuracy (%)" with markers at 0, 20, 40, 60, and 80. * **X-axis:** Categorized into four tasks: "Coding", "Math", "Multiple Choice", and "Creative Writing". * **Legend (Compression Ratio):** * **0%:** Represented by a dashed horizontal line indicating the baseline accuracy for each task. * **50%:** Represented by bars with diagonal hatching (///). * **25%:** Represented by solid bars. * **Legend (Method):** * **REAP (ours) (specific):** Light Blue * **REAP (ours) (general):** Dark Blue * **EAN (specific):** Light Red * **EAN (general):** Dark Red * **Frequency (specific):** Light Green * **Frequency (general):** Dark Green * **HC-SMoE (specific):** Light Yellow/Gold * **HC-SMoE (general):** Dark Yellow/Gold * **M-SMoE (specific):** Light Cyan * **M-SMoE (general):** Dark Cyan ### Data Extraction The following table summarizes the approximate Mean Accuracy (%) for each method, task, and compression ratio, based on the visual representation in the chart. | Task | Method | 0% Compression (Baseline) | 50% Compression (Hatched) | 25% Compression (Solid) | | :--- | :--- | :--- | :--- | :--- | | **Coding** | | **~58%** | | | | | REAP (ours) (specific) | | ~56% | ~54% | | | REAP (ours) (general) | | ~53% | ~51% | | | EAN (specific) | | ~57% | ~55% | | | EAN (general) | | ~32% | ~30% | | | Frequency (specific) | | ~57% | ~55% | | | Frequency (general) | | ~47% | ~45% | | | HC-SMoE (specific) | | ~53% | ~51% | | | HC-SMoE (general) | | ~41% | ~39% | | | M-SMoE (specific) | | ~57% | ~55% | | | M-SMoE (general) | | ~41% | ~39% | | **Math** | | **~89%** | | | | | REAP (ours) (specific) | | ~88% | ~86% | | | REAP (ours) (general) | | ~81% | ~79% | | | EAN (specific) | | ~88% | ~86% | | | EAN (general) | | ~86% | ~84% | | | Frequency (specific) | | ~88% | ~86% | | | Frequency (general) | | ~40% | ~38% | | | HC-SMoE (specific) | | ~85% | ~83% | | | HC-SMoE (general) | | ~73% | ~71% | | | M-SMoE (specific) | | ~89% | ~87% | | | M-SMoE (general) | | ~83% | ~81% | | **Multiple Choice** | | **~72%** | | | | | REAP (ours) (specific) | | ~69% | ~67% | | | REAP (ours) (general) | | ~59% | ~57% | | | EAN (specific) | | ~65% | ~63% | | | EAN (general) | | ~59% | ~57% | | | Frequency (specific) | | ~65% | ~63% | | | Frequency (general) | | ~56% | ~54% | | | HC-SMoE (specific) | | ~67% | ~65% | | | HC-SMoE (general) | | ~54% | ~52% | | | M-SMoE (specific) | | ~56% | ~54% | | | M-SMoE (general) | | ~45% | ~43% | | **Creative Writing** | | **~81%** | | | | | REAP (ours) (specific) | | ~80% | ~78% | | | REAP (ours) (general) | | ~72% | ~70% | | | EAN (specific) | | ~80% | ~78% | | | EAN (general) | | ~70% | ~68% | | | Frequency (specific) | | ~80% | ~78% | | | Frequency (general) | | ~68% | ~66% | | | HC-SMoE (specific) | | ~50% | ~48% | | | HC-SMoE (general) | | ~4% | ~2% | | | M-SMoE (specific) | | ~80% | ~78% | | | M-SMoE (general) | | ~72% | ~70% | ### Key Trends * **General vs. Specific:** Across all tasks and methods, the "specific" version of a method consistently outperforms its "general" counterpart at both 50% and 25% compression ratios. * **Compression Impact:** For all methods and tasks, accuracy decreases as the compression ratio increases from 50% to 25%. * **Baseline Comparison:** In most cases, the compressed models (both 50% and 25%) achieve lower accuracy than the 0% compression baseline. However, some "specific" methods at 50% compression come very close to the baseline, particularly in the "Math" and "Creative Writing" tasks. * **Method Performance:** "REAP (ours) (specific)" and "EAN (specific)" generally show strong performance across tasks, often being the top-performing methods at both compression ratios. "HC-SMoE (general)" shows a significant drop in performance, especially in the "Creative Writing" task.

# Technical Document Extraction: Bar Chart Analysis ## Chart Overview The image is a grouped bar chart comparing **mean accuracy (%)** across four tasks: **Coding**, **Math**, **Multiple Choice**, and **Creative Writing**. The chart evaluates performance under different **compression ratios** (0%, 25%, 50%) and **methods** (REAP, EAN, Frequency, HC-SMoE, M-SMoE), with distinctions between **specific** and **general** implementations. --- ## Key Labels and Axis Titles - **Y-Axis**: "Mean Accuracy (%)" (range: 0–100). - **X-Axis**: Tasks labeled as "Coding", "Math", "Multiple Choice", and "Creative Writing". - **Legend**: - **Compression Ratios**: - 0% (dashed line). - 25% (dotted line). - 50% (solid line). - **Methods**: - **REAP (ours)**: - Specific (solid blue). - General (striped blue). - **EAN**: - Specific (solid red). - General (striped red). - **Frequency**: - Specific (solid green). - General (striped green). - **HC-SMoE**: - Specific (solid yellow). - General (striped yellow). - **M-SMoE**: - Specific (solid cyan). - General (striped cyan). --- ## Data Categories and Sub-Categories 1. **Tasks**: - Coding - Math - Multiple Choice - Creative Writing 2. **Methods**: - REAP (specific/general) - EAN (specific/general) - Frequency (specific/general) - HC-SMoE (specific/general) - M-SMoE (specific/general) 3. **Compression Ratios**: - 0% - 25% - 50% --- ## Key Trends and Data Points 1. **REAP (ours)**: - **Highest accuracy** across all tasks, especially in **Math** (specific: ~90%, general: ~80%) and **Creative Writing** (specific: ~80%, general: ~70%). - Outperforms all other methods in **Math** and **Creative Writing**. 2. **Compression Ratio Impact**: - **0% compression** (dashed line) consistently shows the highest accuracy across all methods and tasks. - **50% compression** (solid line) results in the lowest accuracy, with significant drops compared to 0%. 3. **Specific vs. General Methods**: - **Specific implementations** (solid colors) consistently outperform **general implementations** (striped colors) across all tasks. - Example: In **Coding**, REAP (specific: ~55%) > REAP (general: ~50%). 4. **Task-Specific Performance**: - **Math**: REAP (specific) dominates (~90%), followed by EAN (specific: ~85%) and Frequency (specific: ~80%). - **Creative Writing**: REAP (specific: ~80%) leads, with M-SMoE (specific: ~75%) and EAN (specific: ~70%) close behind. - **Coding**: REAP (specific: ~55%) and EAN (specific: ~50%) are top performers. - **Multiple Choice**: REAP (specific: ~65%) and Frequency (specific: ~60%) are leading. --- ## Transcribed Text from Diagram - **Compression Ratio Labels**: - 0% (dashed line). - 25% (dotted line). - 50% (solid line). - **Method Labels**: - REAP (ours) (specific/general). - EAN (specific/general). - Frequency (specific/general). - HC-SMoE (specific/general). - M-SMoE (specific/general). --- ## Data Table Reconstruction | Task | Method | Compression Ratio | Mean Accuracy (%) | |--------------------|----------------------|-------------------|-------------------| | Coding | REAP (specific) | 0% | ~55 | | Coding | REAP (general) | 0% | ~50 | | Coding | EAN (specific) | 0% | ~50 | | Coding | Frequency (specific) | 0% | ~45 | | Math | REAP (specific) | 0% | ~90 | | Math | REAP (general) | 0% | ~80 | | Creative Writing | REAP (specific) | 0% | ~80 | | Creative Writing | M-SMoE (specific) | 0% | ~75 | | Multiple Choice | REAP (specific) | 0% | ~65 | | Multiple Choice | Frequency (specific) | 0% | ~60 | --- ## Notes - **Visual Estimations**: Accuracy values are approximated from bar heights. - **Pattern Consistency**: Solid colors (specific) > striped colors (general) across all tasks. - **Compression Impact**: Accuracy decreases as compression ratio increases (0% > 25% > 50%).