MedLLAMA-LoRA

Experimental llama finetune on medical qa dataset

This model has not been evaluated yet, and should NOT be used for medical advice. It is an experiment to create a domain-specific model from LLaMA using LoRA finetuning.

Training Details:

13b model, finetuned on 76k question-answer pairs
superset of alpaca-data-cleaned instruct dataset with additional medical qa pairs adapted from icliniq dataset
Trained for 18 hours on A100, minibatch size 10, batch size 256, cutoff_len 512, all other parameters default
https://github.com/tloen/alpaca-lora