|
--- |
|
base_model: |
|
- Qwen/Qwen1.5-72B-Chat |
|
- abacusai/Liberated-Qwen1.5-72B |
|
license: other |
|
license_name: tongyi-qianwen |
|
license_link: https://huggingface.co/Qwen/Qwen1.5-72B-Chat/blob/main/LICENSE |
|
language: |
|
- en |
|
tags: |
|
- mergekit |
|
- merge |
|
- moe |
|
--- |
|
# Qwen1.5-MoE-2x72B |
|
|
|
## Description |
|
This model is created using MoE (Mixture of Experts) through mergekit based on [Qwen/Qwen1.5-72B-Chat](https://huggingface.co/Qwen/Qwen1.5-72B-Chat) and [abacusai/Liberated-Qwen1.5-72B](https://huggingface.co/abacusai/Liberated-Qwen1.5-72B) without further FT. |
|
|
|
It utilizes a customized script for MoE via mergekit, which is available [here](https://github.com/Aratako/mergekit-qwen2). |
|
|
|
Due to the structural modifications introduced by MoE, the use of this model requires [custom modeling file](https://huggingface.co/Aratako/Liberated-Qwen1.5-2x72B/blob/main/modeling_qwen2.py) and [custom configuration file](https://huggingface.co/Aratako/Liberated-Qwen1.5-2x72B/blob/main/configuration_qwen2.py). |
|
When using the model, please place these files in the same folder as the model. |
|
|
|
This model inherits the the [tongyi-qianwen license](https://huggingface.co/Qwen/Qwen1.5-72B-Chat/blob/main/LICENSE). |
|
|
|
## Benchmark |
|
The benchmark score of the [mt-bench](https://github.com/lm-sys/FastChat/tree/main/fastchat/llm_judge) for this model and the two base models are as follows: |
|
|
|
**1-turn, 4-bit quantization** |
|
|Model|Size|Coding|Extraction|Humanities|Math|Reasoning|Roleplay|STEM|Writing|avg_score| |
|
|---|---|---|---|---|---|---|---|---|---|---| |
|
| Liberated-Qwen1.5-72B | 72B | **5.8** | 7.9 | 9.6 | 6.7 | 7.0 | **9.05** | 9.55 | **9.9** | 8.1875 | |
|
| Qwen1.5-72B-Chat | 72B | 5.5 | **8.7** | 9.7 | **8.4** | 7.5 | 9.0 | 9.45 | 9.75 | **8.5000** | |
|
| This model | 2x72B | 5.6 | 7.8 | **9.75** | 7.0 | **8.1** | 9.0 | **9.65** | 9.8 | 8.3375 | |
|
![mt-bench-1turn](./mt-bench-1turn.png) |
|
|
|
**2-turn, 4-bit quantization** |
|
|Model|Size|Coding|Extraction|Humanities|Math|Reasoning|Roleplay|STEM|Writing|avg_score| |
|
|---|---|---|---|---|---|---|---|---|---|---| |
|
| Liberated-Qwen1.5-72B | 72B | 3.9 | 8.2 | **10.0** | 5.7 | 5.5 | 8.4 | 8.7 | 8.6 | 7.3750 | |
|
| Qwen1.5-72B-Chat | 72B | **5.2** | 8.8 | **10.0** | **6.1** | 6.7 | 9.0 | **9.8** | **9.5** | 8.1375 | |
|
| This model | 2x72B | 5.0 | **9.5** | 9.9 | 5.6 | **8.1** | **9.3** | 9.6 | 9.2 | **8.2750** | |
|
![mt-bench-2turn](./mt-bench-2turn.png) |
|
|
|
## Merge config |
|
[mergekit_config.yml](./mergekit_moe_config.yml) |
|
```yaml |
|
base_model: ./Qwen1.5-72B-Chat |
|
gate_mode: random |
|
dtype: bfloat16 |
|
experts: |
|
- source_model: ./Qwen1.5-72B-Chat |
|
positive_prompts: [] |
|
- source_model: ./Liberated-Qwen1.5-72B |
|
positive_prompts: [] |
|
tokenizer_source: model:./Qwen1.5-72B-Chat |
|
``` |
|
|
|
## Gratitude |
|
- Huge thanks to [Alibaba Cloud Qwen](https://www.alibabacloud.com/solutions/generative-ai/qwen) for training and publishing the weights of Qwen model |
|
- Thank you to [abacusai](https://huggingface.co/abacusai) for publishing fine-tuned model from Qwen |
|
- And huge thanks to [mlabonne](https://huggingface.co/mlabonne), as I customized modeling file using [phixtral](https://huggingface.co/mlabonne/phixtral-4x2_8) as a reference |