My base pretrain model has undergone full fine-tuning on an additional 350M tokens using portions of Tulu 3 and Nvidia Nemotron instruct sets.
It is rough but functionsl, and still needs DPO training to align it with human preferences.
** Special note and edit 01/27 - Ministral 3 3B specifically uses tied_word_embeddings set to true in the config.json, which was present in my base fine tune. However after instruct training this model was saved with it set to false, likely due to the trainer confusing it for another model due to the llamaficiation. I have now fixed the flag in this model. If you train on this model be aware that this is something to watch for.
