gpt3_ft.py

import youtube_dl
import openai
import os

# Set up OpenAI credentials
openai.api_key = "YOUR_API_KEY"

# Define function to transcribe audio using OpenAI
def transcribe_audio(audio_file):
    response = openai.Completion.create(
        engine="davinci",
        prompt="Transcribe the following audio:\n" + audio_file,
        max_tokens=2048,
        n=1,
        stop=None,
        temperature=0.5,
    )
    return response.choices[0].text.strip()

# Download YouTube video and extract audio using youtube_dl
url = "https://www.youtube.com/watch?v=YOUR_VIDEO_ID"
ydl_opts = {
    'format': 'bestaudio/best',
    'outtmpl': '%(id)s.%(ext)s',
    'postprocessors': [{
        'key': 'FFmpegExtractAudio',
        'preferredcodec': 'wav',
        'preferredquality': '192',
    }],
}
with youtube_dl.YoutubeDL(ydl_opts) as ydl:
    ydl.download([url])

# Transcribe audio and chop up text into prompts and completions
audio_file = "YOUR_VIDEO_ID.wav"
transcribed_text = transcribe_audio(audio_file)
prompt_length = 50
completion_length = 20
prompts = [transcribed_text[i:i+prompt_length] for i in range(0, len(transcribed_text), prompt_length)]
completions = [transcribed_text[i:i+completion_length] for i in range(prompt_length, len(transcribed_text), completion_length)]

# Fine-tune GPT-3 using prompts and completions
model_engine = "davinci"
model_name = "YOUR_MODEL_NAME"
model_prompt = "\n".join(prompts)
model_completion = "\n".join(completions)
fine_tuned_model = openai.FineTune.create(
    model=model_name,
    prompt=model_prompt,
    examples=[{"text": model_completion}],
    temperature=0.7,
    max_tokens=1024,
    n_epochs=5,
    batch_size=4,
    learning_rate=1e-5,
    labels=["transcription"],
    create= True
)

# Print the fine-tuned model's ID
print(f"Fine-tuned model ID: {fine_tuned_model.id}")