kideng

And command R+

> Do you have some inside knowledge that API based LLMs will support something like this in the future? That would definitely change prompting techniques. While there is no technical...

v0.37.0 does not work either. I tried more early versions and all failed. Maybe I did something wrong? The docker run command I used looks like this: `docker run --rm...

It worked! The devil is in the detail! Thank you!

重启就好了

If you do that, I think you have to maintain two different full-parameter models. But it seems not: `unwrapped_model.pretrained_model.set_adapter("value") _, _, values = model(**input_kwargs) unwrapped_model.pretrained_model.set_adapter("default") logits, _, _ = model(**input_kwargs)`

You are right. I just wonder whether two different LoRA Adapter on one model can be trained simultaneously.

Hey, why is it not updated now, is it found that there will be any bugs?

It seems unnecessary to resolve this issue, because RLOO may perform better and is easier to be supported by llama-factory(trl supports it now).