Please upload base models
We cant train the instruct models because there is significant loss when training instruct models. Please upload the base models also
I think you can just download base models from the other open source repos like https://huggingface.co/meta-llama/Meta-Llama-3-70B ?
@yanjzhou yea but this model is made using a special technique to inference differently with reduced vram. But in order to train with reduced vram, you still need the base model using the eagle technique
I know. I actually trained my eagle model. Did you check the link I put in the comment? You can use that as base models. I don't think the author customized the base model. At least for llama, he used the OSS model as I attached. You can just download the OSS model and train EAGLE. It worked for me.