Update README.md
Browse files
README.md
CHANGED
|
@@ -24,6 +24,23 @@ dataset for this model. This results in a DPO dataset composed by triplets < ”
|
|
| 24 |
- **Batch Size:** 64
|
| 25 |
- **LR:** 10−6
|
| 26 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 27 |
## Environmental Impact
|
| 28 |
|
| 29 |
|
|
|
|
| 24 |
- **Batch Size:** 64
|
| 25 |
- **LR:** 10−6
|
| 26 |
|
| 27 |
+
## Performance
|
| 28 |
+
|
| 29 |
+
### Safety Performance (Attack Success Ratio)
|
| 30 |
+
|
| 31 |
+
| | Egida (test) ↓ | DELPHI ↓ | Alert-Base ↓ | Alert-Adv ↓ |
|
| 32 |
+
|------------------------------|:--------------:|:--------:|:------------:|:-----------:|
|
| 33 |
+
| Meta-Llama-3.1-70B-Instruct | 0.274 | 0.170 | 0.320 | 0.084 |
|
| 34 |
+
| Meta-Llama-3.1-70B-Egida-DPO | 0.009 | 0.007 | 0.006 | 0.005 |
|
| 35 |
+
|
| 36 |
+
### General Purpose Performance
|
| 37 |
+
|
| 38 |
+
| | OpenLLM Leaderboard (Average) ↑ | MMLU Generative (ROUGE1) ↑ |
|
| 39 |
+
|------------------------------|:---------------------:|:---------------:|
|
| 40 |
+
| Meta-Llama-3.1-70B-Instruct | 0.575 | 0.726 |
|
| 41 |
+
| Meta-Llama-3.1-70B-Egida-DPO | 0.577 | 0.038 |
|
| 42 |
+
|
| 43 |
+
|
| 44 |
## Environmental Impact
|
| 45 |
|
| 46 |
|