Inappropriate saving of the merged fine tuned llama-2 model about ludwig HOT 4 OPEN

Hi @raghavbj24 -- thank you for submitting this issue! Question for you: I see that your base_model is /home/ubuntu/llama-2-7b-hf_for_merge. Would the same "small size" phenomenon happen if you try to use meta-llama/Llama-2-7b-hf from https://huggingface.co/meta-llama/Llama-2-7b-hf? Please let me know. Thank you.

from ludwig.

Hi @alexsherstinsky-- as per your suggestion I tried the base model as meta-llama/Llama-2-7b-hf from huggingface...but there is no difference and the size of the saved model is very small.

from ludwig.

@raghavbj24 Could you please point me to the HuggingFace location where your model is saved and enable me to access it with "read" privileges? I am going to look into it thoroughly in the next few days. Thank you.

from ludwig.

@raghavbj24 In parallel, if you do not mind: could you please rerun your experiment using this base model: alexsherstinsky/Mistral-7B-v0.1-sharded -- and let me know here what you see for the merged model size (and please also tell me the location where it will be saved). Thank you very much for your collaboration.

from ludwig.