Update README.md
Browse files
README.md
CHANGED
|
@@ -78,10 +78,10 @@ The datasets for evaluation include:
|
|
| 78 |
|C-Eval | - | - | - | 83.5 | **87.7** |
|
| 79 |
|CMMLU | - | - | 84.8 | 82.3 | **88.5** |
|
| 80 |
| ***Multilingual*** | | | | | |
|
| 81 |
-
|Multi-Exam | - | 56.1 |
|
| 82 |
-
|Multi-Understanding | - | 70.7 |
|
| 83 |
-
|Multi-Mathematics | - | 45.0 |
|
| 84 |
-
|Multi-Translation | - | 29.8 |
|
| 85 |
|
| 86 |
### Efficient MoE Models
|
| 87 |
Compared with training models smaller than 7 billion parameters, it is costly to train medium-size models like 32B while admittedly the 14B model is incapable of performing complex tasks well as the 72B model does. Owing to the recent success of MoE models, this time we turn to employ the MoE model architecture following our previous work Qwen1.5-MoE-A2.7B and extend it to larger model size. Specifically, we apply the same architecture and training strategy, e.g., upcycling, to the model with a total of 57B parameters, only 14B of which are activated in each forward pass. In the following, we list the inference performance of the two models in the deployment with vLLM on 2 NVIDIA A100:
|
|
|
|
| 78 |
|C-Eval | - | - | - | 83.5 | **87.7** |
|
| 79 |
|CMMLU | - | - | 84.8 | 82.3 | **88.5** |
|
| 80 |
| ***Multilingual*** | | | | | |
|
| 81 |
+
|Multi-Exam | - | 56.1 | 58.3 | 61.6 | **65.5** |
|
| 82 |
+
|Multi-Understanding | - | 70.7 | 73.9 | 76.5 | **77.0** |
|
| 83 |
+
|Multi-Mathematics | - | 45.0 | 49.3 | 56.1 | **62.3** |
|
| 84 |
+
|Multi-Translation | - | 29.8 | 30.0 | 33.5 | **34.5** |
|
| 85 |
|
| 86 |
### Efficient MoE Models
|
| 87 |
Compared with training models smaller than 7 billion parameters, it is costly to train medium-size models like 32B while admittedly the 14B model is incapable of performing complex tasks well as the 72B model does. Owing to the recent success of MoE models, this time we turn to employ the MoE model architecture following our previous work Qwen1.5-MoE-A2.7B and extend it to larger model size. Specifically, we apply the same architecture and training strategy, e.g., upcycling, to the model with a total of 57B parameters, only 14B of which are activated in each forward pass. In the following, we list the inference performance of the two models in the deployment with vLLM on 2 NVIDIA A100:
|