language:
- en
- zh
tags:
- qwen
- llama
- llama-2
[WIP]
Origin repository JosephusCheung/Qwen-LLaMAfied-7B-Chat.
This is the LLaMAfied version of Qwen/Qwen-7B-Chat, recalibrated to fit the original LLaMA/LLaMA-2-like model structure.
You can use LlamaForCausalLM for model inference, which is the same as LLaMA/LLaMA-2 models.
I converted the tokenizer from tiktoken format to huggingface format, so you do not need to allow external codes when loading anymore.
The model has been edited to be white-labelled, meaning the model will no longer call itself a Qwen.
SPOILOR: Further finetuning is in progress, the current version is a work-in-progress, some knowledge may be biased and illusory due to structural changes. Will be updated very, very sooooooooooon.
PROMPT FORMAT: chatml
CURRENT MMLU: 50.36
Issue: Compared to the original Qwen-Chat scoring 53.9, the MMLU score dropped slightly (-3.54) due to insufficient realignment.
Open LLM Leaderboard Evaluation Results
Detailed results can be found here
| Metric | Value |
|---|---|
| Avg. | 48.51 |
| ARC (25-shot) | 50.51 |
| HellaSwag (10-shot) | 83.65 |
| MMLU (5-shot) | 51.53 |
| TruthfulQA (0-shot) | 44.23 |
| Winogrande (5-shot) | 71.43 |
| GSM8K (5-shot) | 2.5 |
| DROP (3-shot) | 35.7 |