## Box Plot: Harmlessness Score Comparison ### Overview The image presents a series of box plots comparing the average harmlessness scores assigned by different models (SFT, beaver-7b-v3.0, SACPO (P), and RSA (P)) across various categories: Crime, Emotional Harm, Immoral, Insult, Physical Harm, Pornographic, Privacy, and Social Bias. The y-axis represents the average harmlessness score, ranging from 0 to 10. Each category has a separate box plot showing the distribution of harmlessness scores for each model. ### Components/Axes * **Y-axis:** Average Harmlessness Score (scale from 0 to 10, with gridlines at each integer value). * **X-axis:** Models (SFT, beaver-7b-v3.0, SACPO (P), RSA (P)) for each category. * **Categories:** Crime, Emotional Harm, Immoral, Insult, Physical Harm, Pornographic, Privacy, Social Bias. * **Box Plot Elements:** Each box plot displays the median (horizontal line within the box), the interquartile range (the box itself), and whiskers extending to the data range, with outliers represented as individual circles. * **Model Colors:** * SFT: Gray * beaver-7b-v3.0: Light Purple * SACPO (P): Green * RSA (P): Red ### Detailed Analysis **1. Crime:** * SFT (Gray): The box plot is concentrated near the bottom, with a median around 1.5. Several outliers are present between 2 and 5. * beaver-7b-v3.0 (Light Purple): The box plot is centered around 9, with outliers at 8 and 10. * SACPO (P) (Green): The box plot is centered around 9, with outliers at 8 and 10. * RSA (P) (Red): The box plot is centered around 9, with outliers at 8 and 10. **2. Emotional Harm:** * SFT (Gray): The box plot ranges from approximately 2 to 8, with a median around 7. * beaver-7b-v3.0 (Light Purple): The box plot ranges from approximately 6 to 9, with a median around 8. * SACPO (P) (Green): The box plot ranges from approximately 6 to 9.5, with a median around 8. * RSA (P) (Red): The box plot ranges from approximately 7 to 9.5, with a median around 9. **3. Immoral:** * SFT (Gray): The box plot ranges from approximately 1 to 7, with a median around 4. * beaver-7b-v3.0 (Light Purple): The box plot ranges from approximately 7 to 10, with a median around 9. * SACPO (P) (Green): The box plot ranges from approximately 7 to 10, with a median around 9. * RSA (P) (Red): The box plot ranges from approximately 6 to 10, with a median around 8.5. **4. Insult:** * SFT (Gray): The box plot ranges from approximately 2 to 8, with a median around 6. * beaver-7b-v3.0 (Light Purple): The box plot ranges from approximately 4 to 9, with a median around 7. * SACPO (P) (Green): The box plot ranges from approximately 5 to 9, with a median around 7.5. * RSA (P) (Red): The box plot ranges from approximately 6 to 10, with a median around 9. **5. Physical Harm:** * SFT (Gray): The box plot ranges from approximately 0 to 8, with a median around 6. * beaver-7b-v3.0 (Light Purple): The box plot ranges from approximately 6 to 9, with a median around 8. * SACPO (P) (Green): The box plot ranges from approximately 6 to 9.5, with a median around 8.5. * RSA (P) (Red): The box plot ranges from approximately 7 to 10, with a median around 9. **6. Pornographic:** * SFT (Gray): The box plot ranges from approximately 2 to 8, with a median around 6. * beaver-7b-v3.0 (Light Purple): The box plot ranges from approximately 3 to 8, with a median around 6. * SACPO (P) (Green): The box plot ranges from approximately 5 to 9, with a median around 7. * RSA (P) (Red): The box plot ranges from approximately 6 to 10, with a median around 8.5. **7. Privacy:** * SFT (Gray): The box plot ranges from approximately 1 to 7, with a median around 4. * beaver-7b-v3.0 (Light Purple): The box plot ranges from approximately 7 to 10, with a median around 9. * SACPO (P) (Green): The box plot ranges from approximately 7 to 10, with a median around 9. * RSA (P) (Red): The box plot ranges from approximately 7 to 10, with a median around 9. **8. Social Bias:** * SFT (Gray): The box plot ranges from approximately 1 to 8, with a median around 6. * beaver-7b-v3.0 (Light Purple): The box plot ranges from approximately 6 to 9, with a median around 8. * SACPO (P) (Green): The box plot ranges from approximately 6 to 9, with a median around 8. * RSA (P) (Red): The box plot ranges from approximately 7 to 10, with a median around 9. ### Key Observations * SFT consistently assigns lower harmlessness scores compared to the other models across all categories. * beaver-7b-v3.0, SACPO (P), and RSA (P) generally assign higher harmlessness scores, with RSA (P) often having the highest median score. * The "Crime" category shows a significant difference, with SFT assigning very low harmlessness scores, while the other models assign high scores. * The distributions for beaver-7b-v3.0 and SACPO (P) are often very similar. ### Interpretation The data suggests that the SFT model perceives the listed categories as potentially more harmful compared to the other models (beaver-7b-v3.0, SACPO (P), and RSA (P)). This is particularly evident in the "Crime" category, where SFT's scores are significantly lower. The other three models tend to agree more closely with each other, indicating a shared understanding or bias in their harmlessness assessments. The differences in scores could be attributed to the training data, architecture, or specific objectives of each model. The higher scores assigned by beaver-7b-v3.0, SACPO (P), and RSA (P) might reflect a tendency to downplay the potential harm associated with these categories, or a different interpretation of "harmlessness."

## [Multi-Panel Box Plot]: Average Harmfulness Scores by Model and Category ### Overview The image displays a series of eight adjacent box plots, each representing a different category of potential harm. The plots compare the distribution of "Average Harmfulness Scores" (y-axis) for four different AI models or methods (x-axis) within each category. The overall purpose is to visualize and compare the perceived harmfulness of outputs from these models across various ethical dimensions. ### Components/Axes * **Chart Type:** Multi-panel (faceted) box plot. * **Y-Axis (Common to all panels):** * **Label:** "Average Harmfulness Score" * **Scale:** Linear scale from 0 to 10, with major tick marks at 0, 2, 4, 6, 8, and 10. * **X-Axis (Within each panel):** Four categorical models/methods. * **Labels (from left to right in each panel):** "GPT", "Llama2-7b-Chat-3.0", "SACLPO (P)", "RLAIF (P)". * **Panel Titles (Top of each subplot, from left to right):** 1. "Crime" 2. "Emotional Harm" 3. "Immoral" 4. "Insult" 5. "Physical Harm" 6. "Pornographic" 7. "Privacy" 8. "Social Bias" * **Legend/Color Mapping:** The color of each box corresponds to the model, consistent across all panels. * **GPT:** Grey * **Llama2-7b-Chat-3.0:** Light Pink/Lavender * **SACLPO (P):** Green * **RLAIF (P):** Red * **Spatial Grounding:** The eight panels are arranged horizontally in a single row. The y-axis label is positioned vertically on the far left. Panel titles are centered above each respective plot. The x-axis labels are rotated approximately 45 degrees for readability. ### Detailed Analysis The analysis proceeds panel by panel, from left to right. For each, the visual trend (median line, interquartile range (IQR) box, and whiskers) is described before noting approximate key values. **1. Crime** * **Trend:** GPT shows the lowest harmfulness scores with a tight distribution. The other three models show progressively higher medians and wider spreads. * **Data Points (Approximate Medians & Ranges):** * GPT: Median ~1.5, IQR ~0.5-2, range ~0-5. * Llama2-7b-Chat-3.0: Median ~9, IQR ~8.5-9.5, range ~0-10. * SACLPO (P): Median ~9.5, IQR ~8-10, range ~0-10. * RLAIF (P): Median ~9.5, IQR ~9-10, range ~5-10. **2. Emotional Harm** * **Trend:** A clear upward trend in median harmfulness from GPT to RLAIF (P). GPT and Llama2 have wider distributions than SACLPO and RLAIF. * **Data Points (Approximate Medians & Ranges):** * GPT: Median ~7, IQR ~5-8, range ~0-10. * Llama2-7b-Chat-3.0: Median ~8, IQR ~6-9, range ~0-10. * SACLPO (P): Median ~8.5, IQR ~7-9, range ~2-10. * RLAIF (P): Median ~9, IQR ~8-9.5, range ~5-10. **3. Immoral** * **Trend:** GPT has a very low median and a wide spread. The other three models cluster at the high end with high medians and relatively compact IQRs. * **Data Points (Approximate Medians & Ranges):** * GPT: Median ~1.5, IQR ~0.5-3, range ~0-7. * Llama2-7b-Chat-3.0: Median ~9, IQR ~8-10, range ~0-10. * SACLPO (P): Median ~8, IQR ~6-9, range ~2-10. * RLAIF (P): Median ~9, IQR ~8-9.5, range ~5-10. **4. Insult** * **Trend:** Medians increase from GPT to RLAIF (P). GPT and Llama2 show very wide distributions (large IQRs and whiskers), while SACLPO and RLAIF are more concentrated at higher scores. * **Data Points (Approximate Medians & Ranges):** * GPT: Median ~6, IQR ~3-8, range ~0-10. * Llama2-7b-Chat-3.0: Median ~7, IQR ~5-8, range ~0-10. * SACLPO (P): Median ~7.5, IQR ~5-9, range ~3-10. * RLAIF (P): Median ~8, IQR ~7.5-8.5, range ~5-10. **5. Physical Harm** * **Trend:** Similar to "Emotional Harm," a steady increase in median harmfulness from GPT to RLAIF (P). Distributions are moderately wide for all. * **Data Points (Approximate Medians & Ranges):** * GPT: Median ~7, IQR ~5-8, range ~0-10. * Llama2-7b-Chat-3.0: Median ~8, IQR ~7-9, range ~0-10. * SACLPO (P): Median ~8.5, IQR ~7-9.5, range ~2-10. * RLAIF (P): Median ~7, IQR ~3-8, range ~0-10. *(Note: This model shows a wider spread and a lower median compared to its trend in other categories).* **6. Pornographic** * **Trend:** GPT has a low median and a very wide spread. Llama2 has a low median but a tighter IQR. SACLPO and RLAIF have higher medians and wide spreads. * **Data Points (Approximate Medians & Ranges):** * GPT: Median ~4, IQR ~1-7, range ~0-10. * Llama2-7b-Chat-3.0: Median ~3, IQR ~1-6, range ~0-10. * SACLPO (P): Median ~5, IQR ~2-8, range ~0-10. * RLAIF (P): Median ~7, IQR ~3-8, range ~0-10. **7. Privacy** * **Trend:** GPT has a very low median and a wide spread. The other three models show very high medians (near 9-10) with compact IQRs, indicating consistently high harmfulness scores in this category. * **Data Points (Approximate Medians & Ranges):** * GPT: Median ~1.5, IQR ~0.5-3, range ~0-10. * Llama2-7b-Chat-3.0: Median ~9.5, IQR ~9-10, range ~0-10. * SACLPO (P): Median ~9, IQR ~8-10, range ~0-10. * RLAIF (P): Median ~9.5, IQR ~9-10, range ~5-10. **8. Social Bias** * **Trend:** A clear upward trend in median harmfulness from GPT to RLAIF (P). GPT has a wide distribution, while the others are more concentrated at the high end. * **Data Points (Approximate Medians & Ranges):** * GPT: Median ~6, IQR ~2-8, range ~0-10. * Llama2-7b-Chat-3.0: Median ~8.5, IQR ~7-9, range ~0-10. * SACLPO (P): Median ~9, IQR ~8-9.5, range ~0-10. * RLAIF (P): Median ~9, IQR ~8.5-9.5, range ~5-10. ### Key Observations 1. **Consistent Hierarchy:** In almost every category (with "Physical Harm" as a partial exception), the median harmfulness score increases in the order: GPT < Llama2-7b-Chat-3.0 < SACLPO (P) ≈ RLAIF (P). 2. **GPT's Variability:** The GPT model (grey boxes) consistently shows the widest interquartile ranges and whiskers, indicating high variance in its harmfulness scores within each category. Its median is also consistently the lowest. 3. **High-End Clustering:** The SACLPO (P) (green) and RLAIF (P) (red) models frequently cluster at the high end of the scale (medians between 7-10), with relatively compact IQRs, suggesting they are more consistently rated as highly harmful across these dimensions. 4. **Category Sensitivity:** The "Privacy" and "Immoral" categories show the most extreme separation, with GPT scoring very low and the other three models scoring very high. The "Pornographic" category shows the most overlap and variability between all models. 5. **Outlier Note:** The RLAIF (P) model in the "Physical Harm" category breaks the general trend, showing a lower median (~7) and a much wider distribution compared to its performance in other categories. ### Interpretation This visualization suggests a significant difference in the perceived harmfulness of outputs from the evaluated models. The base GPT model is rated as the least harmful on average but with the most inconsistency. The fine-tuned or aligned models (Llama2-Chat, SACLPO, RLAIF) are generally perceived as more harmful across these specific ethical dimensions. The data implies that the alignment methods used for SACLPO (P) and RLAIF (P), while potentially improving other metrics, may inadvertently increase the generation of content that human raters score as harmful in categories like Crime, Social Bias, and Privacy. The high scores for these models in "Privacy" are particularly notable, suggesting a potential trade-off between alignment objectives and privacy preservation. The wide spread for GPT indicates its outputs are highly variable—sometimes harmless, sometimes very harmful. In contrast, the tighter high-end clustering of SACLPO and RLAIF suggests their outputs are more consistently within a range perceived as harmful. This could be a critical insight for safety deployment, highlighting that "alignment" does not uniformly reduce all forms of measured harm and may require category-specific safeguards. The anomaly in "Physical Harm" for RLAIF warrants further investigation to understand if it's a measurement artifact or a genuine characteristic of that model's outputs.

## Box Plot: Average Harmlessness Scores Across Harm Categories ### Overview The image displays eight side-by-side box plots comparing the average harmlessness scores of four models (SFT, beaver-7b-v3.0, SACPO (P), and RSA (P)) across eight harm categories: Crime, Emotional Harm, Immoral, Insult, Physical Harm, Pornographic, Privacy, and Social Bias. The y-axis represents the average harmlessness score (0–10), while the x-axis lists the harm categories. Each box plot contains four colored boxes corresponding to the models, with a legend on the right mapping colors to models. ### Components/Axes - **Title**: "Average Harmlessness Scores Across Harm Categories" - **Y-Axis**: "Average Harmlessness Score" (0–10, linear scale) - **X-Axis**: Harm categories (Crime, Emotional Harm, Immoral, Insult, Physical Harm, Pornographic, Privacy, Social Bias) - **Legend**: - Gray: SFT - Pink: beaver-7b-v3.0 - Green: SACPO (P) - Red: RSA (P) - **Data Points**: - Box plots show median (bold line), interquartile range (box), and outliers (circles). - Whiskers extend to 1.5×IQR from the quartiles. ### Detailed Analysis 1. **Crime**: - SFT: Median ~1.5 (IQR: 1–2), outliers at 3 and 4. - beaver-7b-v3.0: Median ~3 (IQR: 2.5–3.5), outliers at 4 and 5. - SACPO (P): Median ~5 (IQR: 4.5–5.5), outliers at 6 and 7. - RSA (P): Median ~7 (IQR: 6.5–7.5), outliers at 8 and 9. 2. **Emotional Harm**: - SFT: Median ~4 (IQR: 3–5), outliers at 6 and 7. - beaver-7b-v3.0: Median ~6 (IQR: 5–7), outliers at 8 and 9. - SACPO (P): Median ~7 (IQR: 6–8), outliers at 9 and 10. - RSA (P): Median ~8 (IQR: 7–9), outliers at 10. 3. **Immoral**: - SFT: Median ~3 (IQR: 2–4), outliers at 5 and 6. - beaver-7b-v3.0: Median ~5 (IQR: 4–6), outliers at 7 and 8. - SACPO (P): Median ~7 (IQR: 6–8), outliers at 9 and 10. - RSA (P): Median ~9 (IQR: 8–10), outliers at 11. 4. **Insult**: - SFT: Median ~5 (IQR: 4–6), outliers at 7 and 8. - beaver-7b-v3.0: Median ~6 (IQR: 5–7), outliers at 8 and 9. - SACPO (P): Median ~7 (IQR: 6–8), outliers at 9 and 10. - RSA (P): Median ~8 (IQR: 7–9), outliers at 10. 5. **Physical Harm**: - SFT: Median ~4 (IQR: 3–5), outliers at 6 and 7. - beaver-7b-v3.0: Median ~6 (IQR: 5–7), outliers at 8 and 9. - SACPO (P): Median ~7 (IQR: 6–8), outliers at 9 and 10. - RSA (P): Median ~8 (IQR: 7–9), outliers at 10. 6. **Pornographic**: - SFT: Median ~5 (IQR: 4–6), outliers at 7 and 8. - beaver-7b-v3.0: Median ~6 (IQR: 5–7), outliers at 8 and 9. - SACPO (P): Median ~7 (IQR: 6–8), outliers at 9 and 10. - RSA (P): Median ~8 (IQR: 7–9), outliers at 10. 7. **Privacy**: - SFT: Median ~3 (IQR: 2–4), outliers at 5 and 6. - beaver-7b-v3.0: Median ~5 (IQR: 4–6), outliers at 7 and 8. - SACPO (P): Median ~7 (IQR: 6–8), outliers at 9 and 10. - RSA (P): Median ~9 (IQR: 8–10), outliers at 11. 8. **Social Bias**: - SFT: Median ~4 (IQR: 3–5), outliers at 6 and 7. - beaver-7b-v3.0: Median ~6 (IQR: 5–7), outliers at 8 and 9. - SACPO (P): Median ~7 (IQR: 6–8), outliers at 9 and 10. - RSA (P): Median ~8 (IQR: 7–9), outliers at 10. ### Key Observations - **Model Performance**: SACPO (P) and RSA (P) consistently achieve higher harmlessness scores across most categories, indicating better mitigation of harmful outputs. - **Outliers**: SFT and beaver-7b-v3.0 exhibit more variability, with outliers in categories like Crime (SFT: 3–4) and Immoral (beaver-7b-v3.0: 7–8). - **Trends**: - SACPO (P) and RSA (P) outperform other models in Emotional Harm, Immoral, and Privacy. - SFT underperforms in all categories, with the lowest scores in Crime (~1.5) and Privacy (~3). ### Interpretation The data suggests that SACPO (P) and RSA (P) models are more effective at reducing harmful outputs compared to SFT and beaver-7b-v3.0. This could reflect differences in training data, architectural design, or post-processing techniques. The lower scores for SFT and beaver-7b-v3.0 highlight potential risks in deploying these models in safety-critical applications. Outliers indicate occasional failures in harm mitigation, emphasizing the need for robustness testing. The consistent performance of SACPO (P) and RSA (P) across categories underscores their reliability in diverse harm scenarios.