We record the diagnosis processes of D-Bot(GPT-4), and fine-tune LLMs to simulate the corresponding D-Bot(GPT-4) response (after cleaned). We fine-tune three localized LLMs, i.e., Llama 2-13B, CodeLlama-13B, Baichuan2-13B.
A demo of comparing diagnosis performance of D-Bot(GPT-4) and our Localized D-Bot:
(optional) Step 1: Download Pretrained Model Parameters:
- diag-baichuan2, fine-tuned from Baichuan2-13B
- diag-baichuan2-4bit, a 4bit version of diag-baichuan2
- diag-codellama, fine-tuned from CodeLlama-13B
- diag-llama2, fine-tuned from Llama2-13B
You may also require to replace the ``load'' argument in inference.py with your download path, e.g.,
class DiagBaichuan2Args(BaseModel):
load: str = Field(default="curtis-sun/diag-baichuan2")Step 2: Configure Agents Equipped with localized LLM, e.g., rename config_diag-baichuan2-4bit.yaml as config.yaml.
Then run the project the same as with OpenAI APIs.