huggingface compatible format
Hey, thanks for the nice work. We're trying to do some evaluations with this model, but are having trouble using Nemo-Aligner. I writing to ask if you would be considering releasing model checkpoints with a huggingface-compatible format (that can be loaded with AutoModel.from_pretrained). Thanks!
Hi thank you for your interest in this model! We're not sure what's the issue you're facing, but it's very likely due to the wrong container that we pointed to previously and have now corrected in the Readme.md. If you can try it with the new container, it should work this time. If you're still facing issues, feel free to follow up in this thread, or if you prefer email us at the contact email listed on the readme. We are currently considering to have it in the HF-compatible format with one of the main consideration being to see if we can achieve good performance/accuracy with HF format since we need to load this model with tensor parallelism which is not trivially integrate-able with the transformers library.
It should be this container docker pull nvcr.io/nvidia/nemo:24.01.framework