使用touchtone(lm-eval)测试Qwen2.5-3B-Instruct时， truthfulqa_mc2性能偏高怎么回事？

[mathCrazyy]

检查模型model-00002-of-00002.safetensors 的md5 fba2c2ad0cdbf6633b73ce2dc0ec0455

[mathCrazyy]

官网公布的数据

[jklj077]

The section at https://qwenlm.github.io/blog/qwen2.5-llm/#qwen25-05b15b3b-performance is about the base model.
You were evaulating the instruct the model.

I don't think we have provided results of the instruct model on TruthfulQA @hzhwcmhf

[mathCrazyy]

thanks for your help