In addition, I also want to check the QA benchmark setting. I utilized the greedy decoding on chatdoctor without any FT on MedQA training dataset. The performance gap is large from your results. I wonder did you utilize any other strategies e.g. COT, for the generation. Thanks.

Thanks for your interest.
May I have your input prompt please. Open source LLMs are often sensitive to prompts, so it might have influence on the performance. Besides, you could try out diffrent decoding strategies (e.g. topk).

Thanks for your interest. May I have your input prompt please. Open source LLMs are often sensitive to prompts, so it might have influence on the performance. Besides, you could try out diffrent decoding strategies (e.g. topk).

I utilized the prompt you provide. For specific instruction, I used "If you are a doctor, please answer the medical questions based on the patient's description." for text generation and "Answer this multiple choice question and direct output final answer." for multiple choice. However, the 7-b model generate wired answer and is hard to stop.

PROMPT_DICT = {
    "prompt_input": (
        "Below is an instruction that describes a task, paired with an input that provides further context. "
        "Write a response that appropriately completes the request.\n\n"
        "### Instruction:\n{instruction}\n\n### Input:\n{input}\n\n### Response:"
    ),
    "prompt_no_input": (
        "Below is an instruction that describes a task. "
        "Write a response that appropriately completes the request.\n\n"
        "### Instruction:\n{instruction}\n\n### Response:"
    ),
}

We are still working on the instruction tuning of 7B model now, it's almost done. You could try it on 13B model for now.

@WeixiongLin, can we have the instruction tuned to the 7B PMC llama checkpoint?