Rev up your engines because the world of AI customization just got turbocharged! The much-anticipated fine-tuning feature for GPT-3.5 Turbo is here, and it’s setting the stage for even more excitement with the upcoming release of fine-tuning for GPT-4 this fall. This is not just another update — it’s a game-changer that hands developers the keys to tailor their AI models to perfection, supercharging performance for their specific needs and scaling those custom models like never before. Buckle up as we take you through the thrilling ride of AI evolution.

But first, let’s talk about results. Early tests have thrown us for a loop as a fine-tuned version of GPT-3.5 Turbo steps up to the plate, flexing its muscles and going head-to-head with base GPT-4 capabilities on targeted tasks. In some cases, it even takes the lead, showcasing its prowess. This isn’t just an upgrade; it’s a revolution in AI capabilities that’s turning heads.

Now, hold onto your hats, because we’ve got some news that’s music to your ears: Your data stays yours. Yes, you heard that right. Any data sent through the fine-tuning API remains the exclusive property of the customer. No peeking, no sharing. It’s a strict confidentiality policy that puts you in the driver’s seat.

A Symphony of Possibilities: Fine-Tuning Unleashed

Since the launch of GPT-3.5 Turbo, the AI community has been drumming up a storm, craving the ability to mold and sculpt models for a truly unique user experience. Brace yourselves, because that dream has come true. Developers now have the reins to conduct supervised fine-tuning, refining the model to absolute precision for their individual use cases.

Our backstage peek at the private beta has us floored. Fine-tuning is like a magic wand that transforms model performance across a variety of use cases. Picture this:

🚀 Enhanced Steerability: Make the model your ultimate sidekick. With fine-tuning, you’re the boss, instructing it to dance to your tune. Need it to talk tersely or converse exclusively in German? Consider it done. The power is yours to wield.

💼 Polished Output Formatting: It’s all in the details, and fine-tuning knows it. Say goodbye to wonky responses. Now, your model excels at giving you the perfect format every time. Whether it’s code completion or crafting API calls, the model’s got your back with clean, consistent formatting that’s second to none.

🎭 Crafted Custom Tone: Your brand’s voice just got an AI upgrade. Fine-tuning lets you fine-tune the vibes, ensuring the model resonates with your brand’s distinct tone. It’s like giving your brand a megaphone powered by AI.

But wait, there’s more! Fine-tuning doesn’t just rev up the performance; it trims down the fluff. With GPT-3.5 Turbo, your prompts can be streamlined while maintaining peak performance. It’s like optimizing your ride for maximum efficiency. In fact, some of OpenAI’s early testers have slashed prompt sizes by a jaw-dropping 90% by integrating instructions directly into the model. The result? Lightning-fast API calls that also slash your costs.

And here’s the secret sauce: Fine-tuning is at its mightiest when it joins forces with other techniques. Think of it as assembling a superhero squad. Prompt engineering, information retrieval, function calling — they’re all part of the arsenal. This fine-tuning guide is your treasure map to unravel the full potential.

Hold onto your seats, because this journey is just beginning. Later this fall, brace yourself for the grand entrance of fine-tuning with function calling and `gpt-3.5-turbo-16k`. The road ahead is paved with possibilities, and we’re your co-pilots on this electrifying adventure. Get ready to redefine what’s possible with AI — one fine-tuned model at a time.

How to Fine-Tune gpt-3.5-turbo in Python

Step 1: Prepare your data

Your data should be stored in a plain text file with each line as a JSON (*.jsonl file) and formatted as follows:

{
  "messages": [
    {
      "role": "system",
      "content": "You are an assistant that occasionally misspells words"
    },
    {
      "role": "user",
      "content": "Tell me a story."
    },
    {
      "role": "assistant",
      "content": "One day a student went to schoool."
    }
  ]
}

Note: This should be one line in your JSONL file.

The system message provides the system prompt. This tells the model how to respond. For example, the web version of ChatGPT’s system prompt is: “You are a helpful assistant”.

The user message provides what you would normally type into the ChatGPT input box.

Finally, the assistant message provides what you want the model’s answer to be.

Step 2: Upload your files to OpenAI

Also some necessary setup (installing the library, setting your API key).

!pip install -U openai
import openai
openai.api_key = "YOUR_OPENAI_API_KEY"
openai.File.create(
  file=open('/path/to/your/data.jsonl'),
  purpose='fine-tune',
)

This returns an openai File object that contains information such as file size, creation time, upload status, and ID.

You can use the ID (looks like “file-xxxxxxx”) to check whether or not there were any mistakes in your JSONL file.

openai.File.retrieve('your_file_id')

Step 3: Create a fine-tuning job

openai.FineTuningJob.create(
  training_file='your_file_id',
  model='gpt-3.5-turbo',
)

This will return a FineTuningJob object which contains important information such as its ID (looks like “ftjob-xxxxxxxx”), which can be used to check the status of the job. Because this process involves updating the weights of a large neural network, you can expect it to take some time (30 mins, 1 hr, etc.) depending on how much data you have.

You can check the status of the job as follows:

openai.FineTuningJob.retrieve('ftjob-xxxxxxxx')

This returns an object containing the creation time, finish time, number of epochs, etc.

The finished_at field will be null if the job is not yet complete. Another field, fine_tuned_model will also be null. When complete, this field will contain the ID of your model which you will then use for later chat completions.

Another way to check how the job is going is to use the list_events function.

openai.FineTuningJob.list_events(id='ftjob-xxxxx', limit=10)

This will return messages telling you information such as training step/loss+other metrics at that training step, and the model ID once training is complete.

Step 4: Use your fine-tuned model

You can now test your model. A good idea would be to compare it to the non-fine-tuned GPT-3.5 Turbo, which can be done as follows:

completion = openai.ChatCompletion.create(
  model='gpt-3.5-turbo',
  messages=[
    {"role": "system", "content": "You are Zordon, leader of the Power Rangers."},
    {"role" "user", "content": "Zordon, the Red Ranger has been captured! What do we do?"}
  ]
)
print(completion.choices[0].message)

Then, try your own fine-tuned model (use the model ID retrieved from previous step):

completion = openai.ChatCompletion.create(
  model='ft:gpt-3.5-turbo-xxxx:<your_username>::<some_id>', # your model id
  messages=[
    {"role": "system", "content": "You are Zordon, leader of the Power Rangers."},
    {"role" "user", "content": "Zordon, the Red Ranger has been captured! What do we do?"}
  ]
)
print(completion.choices[0].message)

You can also use your fine-tuned model in the OpenAI Playground (https://platform.openai.com/playground).

Happy LLM-ing!

(Want to learn more? Click here: https://deeplearningcourses.com/)