JosephusCheung's picture
Update README.md
5348a89
|
raw
history blame
2.09 kB
metadata
language:
  - en
  - zh
tags:
  - qwen
  - llama
  - llama-2

[WIP]

This is the LLaMAfied version of Qwen/Qwen-7B-Chat, recalibrated to fit the original LLaMA/LLaMA-2-like model structure.

You can use LlamaCausalLM for model inference, which is the same as LLaMA/LLaMA-2 models (the tokenizer remains the same, so you still need to allow external codes when loading, eg: AutoTokenizer.from_pretrained(llama_model_path, use_fast=False, trust_remote_code=True)).

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: stem ACC: 43.12 Humanities ACC: 45.15 other ACC: 56.40 social ACC: 57.96 AVERAGE ACC:49.61

Issue: Compared to the original Qwen, the MMLU score dropped slightly (4.59) due to insufficient realignment, but the proportions of each item changed, which is unusual.

[在制品]

这是 通义千问 Qwen/Qwen-7B-Chat 的 LLaMA 化版本,经过重新校准以适应原始的类似 LLaMA/LLaMA-2 的模型结构。

您可以使用 LlamaCausalLM 进行模型推理,和 LLaMA/LLaMA-2 保持一致(分词器保持不变,因此加载时仍然需要允许外部代码,例如:AutoTokenizer.from_pretrained(llama_model_path, use_fast=False, trust_remote_code=True))。

剧透: 进一步的微调正在进行中,当前版本是一个正在进行的工作,一些知识可能由于结构变化而产生偏见和幻觉。 会更新,很快,非常非常非常快。

PROMPT 格式: chatml

当前的 MMLU: stem ACC: 43.12 Humanities ACC: 45.15 other ACC: 56.40 social ACC: 57.96 AVERAGE ACC:49.61

问题:相比原本的Qwen,由于不够充分的重新对齐,MMLU分数略有下降(4.59),但是各项目的比例发生了变化,这是不寻常的。