language:
- en
- zh
tags:
- qwen
- llama
- llama-2
[Probeversion]
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 (using GPT2Tokenizer converted from the original tiktoken, by vonjack).
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 soooooooooon.
PROMPT FORMAT: chatml
CURRENT MMLU: 53.48
CURRENT CEval (val): 54.13
MMLU - stem ACC: 46.40 Humanities ACC: 47.61 other ACC: 61.31 social ACC: 61.78 AVERAGE ACC:53.48
CEval (val) - STEM acc: 45.28 Social Science acc: 66.19 Humanities acc: 58.76 Other acc: 54.62 Hard acc:28.64 AVERAGE acc:54.13
Issue: Compared to the original Qwen-7B-Chat scoring 53.90 in MMLU and 54.18 in CEval (val), the our scores dropped slightly [-0.42 in MMLU, -0.05 in CEval (val)] due to insufficient realignment.
[预览版本]
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（使用由 vonjack 从原始 tiktoken 转换而来的 GPT2Tokenizer 分词器）。
模型已经被编辑实现白标化，不再自称通义千问。
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，非常非常快。
PROMPT 格式: chatml
当前的 MMLU: 53.48
MMLU - stem ACC: 46.40 Humanities ACC: 47.61 other ACC: 61.31 social ACC: 61.78 AVERAGE ACC:53.48
CEval (val) - STEM acc: 45.28 Social Science acc: 66.19 Humanities acc: 58.76 Other acc: 54.62 Hard acc:28.64 AVERAGE acc:54.13
问题：相比原本的 Qwen-7B-Chat 的 MMLU 分数 53.90 和 CEval (val) 分数 54.13，由于不够充分的重新对齐，分数都略有下降（MMLU -0.42, CEval -0.05）。