|
--- |
|
license: other |
|
license_name: qwen |
|
license_link: LICENSE |
|
--- |
|
|
|
# 🦙 Qwen-72B-Llama |
|
|
|
This is the 🦙 llamafied version of [Qwen/Qwen-72B](https://huggingface.co/Qwen/Qwen-72B). |
|
|
|
## 🛠️ Reproduction |
|
|
|
I used [this script](https://github.com/hiyouga/LLaMA-Factory/blob/main/tests/llamafy_qwen.py) to convert the weights: |
|
|
|
[LLaMA-Factory/tests/llamafy_qwen.py](https://github.com/hiyouga/LLaMA-Factory/blob/main/tests/llamafy_qwen.py) |
|
|
|
## 🔠 Tokenizer |
|
|
|
After I converted the weights, I took the tokenizer from [KnutJaegersberg/Qwen-14B-Llamafied](https://huggingface.co/KnutJaegersberg/Qwen-14B-Llamafied) and uploaded it to this repository. |
|
|
|
## 📊 Eval Scores Compared to Original Model |
|
|
|
Here are some of the evaluation score comparisons based on the [Open LLM Leaderboard](https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard). |
|
|
|
| Metric | Qwen-72B | **Qwen-72B-Llama** | |
|
|-----------------------|---------------|--------------------| |
|
| Avg. | 73.6 | **69.53** | |
|
| ARC (25-shot) | 65.19 | **64.85** | |
|
| HellaSwag (10-shot) | 85.94 | **83.27** | |
|
| MMLU (5-shot) | 77.37 | **73.66** | |
|
| TruthfulQA (0-shot) | 60.19 | **57.6** | |
|
| Winogrande (5-shot) | 82.48 | **81.53** | |
|
| GSM8K (5-shot) | 70.43 | **56.25** | |