[Question]: 阿里云百炼平台发布的qwen2.5-72B和开源qwen2.5-72B能力差距大

[darvsum]

Has this been raised before?

• I have checked the GitHub README.
• I have checked the Qwen documentation and cannot find an answer there.
• I have searched the issues and there is not a similar one.
• I confirm that this is not a bug report, a feature request, or a badcase.

Description

对比了一下阿里云百炼平台和开源的qwen2.5-72B。使用两个模型分别抽取知识图谱，开源的只能抽取出节点，关系一条都没抽取出来，而百炼平台的模型可以抽取出节点和关系。是不是说明开源的qwen2.5-72B做了简化处理？

[jklj077]

Did you use the instruct models? The models are the same. The difference should mainly came from how you run the model.

[monkeywl2020]

台的模型可以抽取出节点和关系。是不是说明开源的qwen2.5-72B做了简化处理？

我试过，可以抽取知识图谱的关系节点，可能需要修改下提示词，并且修改下返回内容的解析函数就可以了。这部分需要自己适配。

[darvsum]

#>Did you use the instruct models? The models are the same. The difference should mainly came from how you run the model.
两个都是使用的 instruct models，代码也是一样的。 @jklj077

[darvsum]

@monkeywl2020 ，我用的graph_transformers默认的提示词。开源模型换成qwen2-7b-instruct竟然抽取实体比qwen2.5-72b-instruct效果好。但是几个开源模型抽取关系始终是空值。

[jklj077]

langchain? what did you actually use?

[jklj077]

it's the same model. wait did you use ollama?

[jklj077]

This is result from using the OpenAI-compatible API from vLLM. It just works.

[darvsum]

是的，使用了ollama，但还是用其他兼容openai的本地api服务，结果都一样。没有试vllm，我试试。

[darvsum]

@jklj077 经过验证使用vllm可行。感谢！
1、使用的是相同的模型，运行多次，节点抽取数始终就这几个，没有你的多

2、node_properties对应的值本地模型没有，线上有

[darvsum]

我更换模型为qwen2-72b-instruct，达到了满意结果。是不是说明qwen2.5在文本理解方面能力减弱了呢？