YichiZhang

Results 6 comments of YichiZhang

Missing `content` field when `finishReason` is other than `STOP`

I also have the same issue. This makes the query from API very unstable, frequently resulting in null response. :(

Training loss curve like stairs.

> > This is the loss curve I have. @yuanzhedong Hi, is this kind of loss curve normal or expected ? I am reproducing the finetuning of llama-7b and want...

The channel to access the data seems to be closed

We have released our data on huggingface, which will be more convenient for you to get data. We are closing this issue.

Improve discoverability of your work on Hugging Face

Thanks for your advice! We have put these on our agenda : )

Code for generating Adversarial examples (untargeted/targeted)

Sure, the corresponding code for online generation has been provided in https://github.com/thu-ml/MMTrustEval/blob/main/mmte/methods/adv.py. We will soon release an example for how to integrate it into the evaluation pipeline.

Regarding testing my own model

Sure. You can integrate your own model to the framework following these steps. 1. Define a subclass of `BaseChat` for your own model 2. Implement the `chat` method to support...