The config.json does not seem to match the new instruction finetuning.
The config.json seems to still be the default with vocab_size of 32000, and the old bos_token_id and eos_token_id. This doesn't match the new tokens added in added_tokens.json and specified in the github.
Ok, I am noticing now that the embedding weights are only of size 32000. I assume this means the model was not finetuned with the new vocabulary? Is the model that is uploaded the instruction finetuned model mentioned on the github or something else? https://github.com/epfLLM/meditron#downstream-use
Same question encountered when using 7B model
Hi there, thank you for bringing this to our attention.
Here is a related issue with our reply:
https://huggingface.co/epfl-llm/meditron-7b/discussions/5
Let us know if this resolves the issue. Looking forward to your feedback!