Constitutional AI models do not achieve MT-Bench scores as reported

Hi, thanks for your great work!

I'm especially interested in the recently-introduced constitutional-ai tuning in this [blog post](https://huggingface.co/blog/constitutional_ai). I've found the open-source [SFT model](https://huggingface.co/alignment-handbook/mistral-7b-sft-constitutional-ai) and [DPO model](https://huggingface.co/HuggingFaceH4/mistral-7b-anthropic) on huggingface. However, when I tried to launch the MT-Bench test with them, the returned results are significantly worse than those reported in the blog post, according to the figure below (which I've copied over here for reference):


The MT-Bench score I've collected are 5.33 / 6.39 for the SFT / DPO model respectively, where the reference figure shows (approximately) 6.5 / 7.2.