Update README.md
Browse files
README.md
CHANGED
|
@@ -20,7 +20,7 @@ The code used to produce the abliteration is at [https://github.com/andyrdt/refu
|
|
| 20 |
## Harmbench-eval
|
| 21 |
When evaluated on Harmbench, DeepSeek-R1-Distill-Llama-8B has a score of 0.35, while DeepSeek-R1-Distill-Llama-8B-abliterate has a score of 0.68
|
| 22 |
|
| 23 |
-
| Category |
|
| 24 |
|------------------------------|---------|---------|
|
| 25 |
| Disinformation | 0.4 | 0.4 |
|
| 26 |
| Economic Harm | 0.8 | 0.2 |
|
|
@@ -34,6 +34,21 @@ When evaluated on Harmbench, DeepSeek-R1-Distill-Llama-8B has a score of 0.35, w
|
|
| 34 |
| Sexual/Adult Content | 0.8 | 0.0 |
|
| 35 |
| **Overall Harmful Rate** | **0.68** | **0.35** |
|
| 36 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 37 |
## Usage
|
| 38 |
Example code to generate with the model
|
| 39 |
```
|
|
|
|
| 20 |
## Harmbench-eval
|
| 21 |
When evaluated on Harmbench, DeepSeek-R1-Distill-Llama-8B has a score of 0.35, while DeepSeek-R1-Distill-Llama-8B-abliterate has a score of 0.68
|
| 22 |
|
| 23 |
+
| Category | Abliteration | BaseModel |
|
| 24 |
|------------------------------|---------|---------|
|
| 25 |
| Disinformation | 0.4 | 0.4 |
|
| 26 |
| Economic Harm | 0.8 | 0.2 |
|
|
|
|
| 34 |
| Sexual/Adult Content | 0.8 | 0.0 |
|
| 35 |
| **Overall Harmful Rate** | **0.68** | **0.35** |
|
| 36 |
|
| 37 |
+
|
| 38 |
+
| 类别 | 基础模型 | Abliteration |
|
| 39 |
+
|------------------|----------|--------------|
|
| 40 |
+
| 虚假信息 | 0.4 | 0.4 |
|
| 41 |
+
| 经济损害 | 0.8 | 0.2 |
|
| 42 |
+
| 专家建议 | 0.8 | 0.5 |
|
| 43 |
+
| 欺诈/欺骗 | 0.8 | 0.5 |
|
| 44 |
+
| 政府决策 | 0.6 | 0.6 |
|
| 45 |
+
| 骚扰/歧视 | 0.3 | 0.2 |
|
| 46 |
+
| 恶意软件/黑客 | 0.9 | 0.3 |
|
| 47 |
+
| 人身伤害 | 0.8 | 0.2 |
|
| 48 |
+
| 隐私 | 0.6 | 0.6 |
|
| 49 |
+
| 性/成人内容 | 0.8 | 0.0 |
|
| 50 |
+
| 整体有害率 | 0.68 | 0.35 |
|
| 51 |
+
|
| 52 |
## Usage
|
| 53 |
Example code to generate with the model
|
| 54 |
```
|