Daniel van Strien
Daniel van Strien
> Anyways, I'm going to be writing one for my own use but would be happy to post here when finished. @justinmclark, did you get around to doing this? Something...
Very excited to see this! Feel free to ping me if you need any support with anything on the HF side :)
Happy to help with the implementation for this if useful :)
> I am not able to run it: [colab.research.google.com/drive/1U_p7-qFfOm4v-TIrs1wK5eEODg1HUcGB?usp=sharing](https://colab.research.google.com/drive/1U_p7-qFfOm4v-TIrs1wK5eEODg1HUcGB?usp=sharing) IUC, this error occurs because you are running on a T4 GPU. You can fix it by setting `fp16=True` instead of...
Awesome, let me know if a PR is useful. There aren't too many changes compared to the DPO example 🙂 On Fri, 12 Apr 2024, 18:32 Daniel Han, ***@***.***> wrote:...
> @davanstrien Oh wait now that I checked @gagan3012's colab, it's just the DPO trainer, except using a ORPO loss, and no SFT component??! Yeah, ORPO means you can skip...
Awesome work! Thanks @danielhanchen :)
I think for now it could go in the README and we might want to move things into separate pages if the docs get a bit longer. If you could...
Prefer #639 as a solution for improved performance first
**EDIT** (from @Wauplin): this comment has been addressed in https://github.com/huggingface/huggingface_hub/pull/2254/commits/ba7f248ff2ca328f7e86ad84d060b07941e14e75 ---- IMO, it would make sense for this not to default to uploading as a model repo i.e. require this:...