data-cleaning-llm / app /openai_chat_completion.py
cmagganas's picture
Upload folder using huggingface_hub
dc52018
raw
history blame
4.13 kB
import os
from io import BytesIO
import pandas as pd
from dotenv import load_dotenv
load_dotenv()
import openai
import streamlit as st
# # set OPENAI_API_KEY environment variable from .streamlit/secrets.toml file
openai.api_key = st.secrets["OPENAI_API_KEY"]
# # set OPENAI_API_KEY environment variable from .env file
# openai.api_key = os.getenv("OPENAI_API_KEY")
# # read in llm-data-cleaner/prompts/gpt4-system-message.txt file into variable system_message
# system_message = open('../prompts/gpt4-system-message.txt', 'r').read()
class OpenAIChatCompletions:
def __init__(self, model="gpt-4", system_message=None):
self.model = model
self.system_message = system_message
# function to input args such as model, prompt, etc. and return completion
def openai_chat_completion(self, prompt, n_shot=None):
messages = [{"role": "system", "content": self.system_message}] if self.system_message else []
# add n_shot number of samples to messages list ... if n_shot is None, then only system_message and prompt will be added to messages list
if n_shot is not None:
messages = self._add_samples(messages, n_samples=n_shot)
messages.append({"role": "user", "content": prompt})
# set up the API request parameters for OpenAI
chat_request_kwargs = dict(
model=self.model,
messages=messages,
)
# make the API request to OpenAI
response = openai.ChatCompletion.create(**chat_request_kwargs)
# return only the completion text
# return response['choices'][0]['message']['content']
# return response
return response
# function to use test data to predict completions
def predict_jsonl(
self,
path_or_buf='../data/cookies_train.jsonl',
# path_or_buf='~/data/cookies_train.jsonl',
n_samples=None,
n_shot=None
):
jsonObj = pd.read_json(path_or_buf=path_or_buf, lines=True)
if n_samples is not None:
jsonObj = jsonObj.sample(n_samples, random_state=42)
iter_range = range(len(jsonObj))
prompts = [jsonObj.iloc[i]['prompt'] for i in iter_range]
completions = [jsonObj.iloc[i]['completion'] for i in iter_range]
predictions = [self.openai_chat_completion(prompt, n_shot=n_shot) for prompt in prompts]
return prompts, completions, predictions
# a method that adds prompt and completion samples to messages
@staticmethod
def _add_samples(messages, n_samples=None):
if n_samples is None:
return messages
samples = OpenAIChatCompletions._sample_jsonl(n_samples=n_samples)
for i in range(n_samples):
messages.append({"role": "user", "content": samples.iloc[i]['prompt']})
messages.append({"role": "assistant", "content": samples.iloc[i]['completion']})
return messages
# a method that samples n rows from a jsonl file, returning a pandas dataframe
@staticmethod
def _sample_jsonl(
path_or_buf='data/cookies_train.jsonl',
# path_or_buf='~/data/cookies_train.jsonl',
n_samples=5
):
# jsonObj = pd.read_json(path_or_buf=path_or_buf, lines=True)
# if running locally, True
# else running on HF Spaces, False
if "Kaleidoscope Data" in os.getcwd():
# file_path = os.path.join(os.getcwd(), "..", path_or_buf)
file_path = os.path.join("/".join(os.getcwd().split('/')[:-1]), path_or_buf)
else:
file_path = os.path.join(os.getcwd(), path_or_buf)
try:
with open(file_path, "r") as file:
jsonl_str = file.read()
jsonObj = pd.read_json(BytesIO(jsonl_str.encode()), lines=True, engine="pyarrow")
except FileNotFoundError:
# Handle the case where the file is not found
# Display an error message or take appropriate action
st.write(f"File not found: {file_path}")
return jsonObj.sample(n_samples, random_state=42)