[NeurIPS 2024] Discovery of the Hidden World with Large Language Models

This repository contains the source codes for reproducing the results of NeurIPS'24 paper: Discovery of the Hidden World with Large Language Models.

Author List: Chenxi Liu*, Yongqiang Chen*, Tongliang Liu, Mingming Gong, James Cheng, Bo Han, Kun Zhang

(* Equal Contribution)

1. Go through the OpenAI Quickstart Tutorial
2. In this project, we LLM_openai() to contorl the interaction with OpenAI

   import sys
   sys.path.insert(0, '/project/root/path') # set project root path
   
   from utils import Logger
   from openai_utils import LLM_openai
   
   # check the conversation log at 'logs/{exp_name}'
   logger = Logger(
       root = '/project/root/path', 
       exp_name = 'set_exp_name'
   )
   
   llm = LLM_openai(logger)
   
   response = llm.chat(
       model="gpt-4-1106-preview",
       content= "Hello GPT4!",
       system_instruct = "You are a helpful assistant."
   )

1. Go through the Poe FastAPI Tutorial
2. In this project, we chat() to contorl the interaction with Poe

   import fastapi_poe as fp
   import asyncio
   
   async def chat(content, bot_name="GPT-3.5-Turbo" ):
       poe_api_key = "your api key" # set your api key
       message = fp.ProtocolMessage(role="user", content=content)
       responce = ""
       async for partial in fp.get_bot_response(messages=[message], bot_name=bot_name, api_key=poe_api_key): 
           responce = responce + partial.text
   
       return responce
   
   response = await chat("Hello world!", bot_name="GPT-3.5-Turbo" )

• Apple_Gastronome_AG7_v20240513.ipynb This is the notebook to generate the AppleGastronome Benchmark dataset. "AG7" means there are 7 variables considered in the data-generating process.
• Apple_Gastronome_AG7_v20240513.xlsx This is the generated AppleGastronome Benchmark dataset. It consists of 8 columns: 7 variables plus one column for the review texts.
• Formal_exp_1 AG7_2024_05_16_1422 GPT-3.5-Turbo release.ipynb This is the notebook to test COAT method with different Foundation Models.
• Other .xlsx files Annotations generated by annotation-2024_05_16_1422_hk_mistralai_GPT-3.5-Turbo release.ipynb

Conversation Logs:

• META: https://poe.com/s/Suv0XqninMg9nQTGFEiy
• COAT+CoT: https://poe.com/s/x1tha9wV6opPydhym2sQ
• DATA: First round of COAT.
• COAT: https://poe.com/s/tYvcQuIyRB54wVpyUZyx

Source of the causal graph: here

• SimulatedData_SampleSize100.csv and neuro_addition_1000_0126.csv Generated from Neuropathic Pain Diagnosis Simulator. Simulation of the real-life diagnosis with three levels of variables, including the symptom-level, radiculopathy-level and the pathophysiology-level.
• id_name.txt From Neuropathic Pain Diagnosis Simulator. It stores variable names.
• Neuropathic_Pain_Diagnosis_v20240119.ipynb This is the notebook to generate clinical notes from tabular data SimulatedData_SampleSize100.csv. Only symptom-level vairables are used.
• neuro_R_shoulder_impingement.xlsx This is the generated clinical notes.
• Formal_exp_1_Neuro_RSI symptom_only release.ipynb This is the notebook to test COAT method with different Foundation Models.