FAQ

Question 1

How does a small amount of fine tuning data create enough impact on model parameters that have been tuned using huge amounts of data?

Accepted Answer

Fine-tuning is a process that allows you to customize a model for your specific application. It is a powerful technique that can yield higher quality results than prompt design, and it enables the model to train on more examples than can fit in a prompt. This results in token savings due to shorter prompts and lower latency requests.

The fine-tuning process involves the following steps:

Prepare and upload training data: Your data must be a JSONL document, where each line is a prompt-completion pair corresponding to a training example. The more training examples you have, the better. It is recommended to have at least a couple hundred examples.
Train a new fine-tuned model: You can start your fine-tuning job using the OpenAI CLI. The choice of base model influences both the performance of the model and the cost of running your fine-tuned model. Your model can be one of: ada, babbage, curie, or davinci.
Use your fine-tuned model: When a job has succeeded, the fine_tuned_model field will be populated with the name of the model. You may now specify this model as a parameter to the Completions API, and make requests to it using the Playground.

Fine-tuning is currently only available for the following base models: davinci, curie, babbage, and ada. These are the original models that do not have any instruction following training. You are also able to continue fine-tuning a fine-tuned model to add additional data without having to start from scratch.

Now, to answer your question, how does a small amount of fine-tuning data create enough impact on model parameters that have been tuned using huge amounts of data?

The key lies in the fact that the base models (like GPT-3) have already been trained on a vast amount of data, learning a wide range of language patterns, structures, and concepts. This pre-training phase allows the model to understand language at a general level. However, these models might not be specialized for specific tasks or domains.

When we fine-tune these models on a smaller, task-specific dataset, we are essentially nudging the model parameters slightly to adapt to the specific task or domain. Even though the fine-tuning data is much smaller compared to the original training data, it is highly relevant to the task at hand, and thus, these small adjustments can lead to significant improvements in performance on the specific task.

It’s important to note that fine-tuning needs to be done carefully. If the fine-tuning data is too different from the data the model was originally trained on, or if the model is fine-tuned too aggressively, it can lead to a problem known as “catastrophic forgetting,” where the model forgets the general language understanding it learned during pre-training. Therefore, it’s a balance between retaining the general language understanding from the pre-training phase and adapting to the specific task during the fine-tuning phase.

Question 2

Would in-context training be able to do customer support type training equally as well as fine-tuning, or in your experience does fine-tuning work better? I only ask because in-context training is significantly cheaper and easier but I'm trying to determine which use case lends itself better to fine-tuning vs in-context training?

Accepted Answer

In the context of AI models, both in-context learning and fine-tuning are valuable techniques for adapting a model to specific tasks. The choice between the two largely depends on the nature of the task, the resources available, and the desired level of customization.

In-context learning is a method where the model is given a series of instructions or examples within the prompt itself. This is a more straightforward and cost-effective approach, as it doesn’t require additional training of the model. It’s essentially “on-the-fly” training, where the model uses the provided context to generate a response.

In the context of customer support, in-context learning can be quite effective. For example, you can provide a few examples of how you want the model to respond to certain queries at the beginning of the prompt, and the model will attempt to follow the pattern in its responses. This can be particularly useful for handling common queries or providing generic responses.

However, in-context learning has its limitations. The model’s responses are only as good as the examples provided, and it may not always generalize well to unseen queries or complex situations. Furthermore, the amount of context that can be provided is limited by the model’s token limit.

Fine-tuning, on the other hand, involves training the model on a specific dataset to adapt it to a particular task. This is a more involved process and requires more resources, as it involves additional training of the model. However, it can lead to more accurate and reliable performance, especially for complex tasks.

In the case of customer support, fine-tuning can be beneficial if you have a large dataset of past customer interactions and responses. By fine-tuning the model on this data, you can train it to handle a wide range of queries and situations and to respond in a manner consistent with your organization’s customer service standards.

Fine-tuning can also be useful for tasks that require a deep understanding of a specific domain or specialized knowledge. For example, if your customer support involves technical products or services, fine-tuning the model on relevant data can help it provide more accurate and helpful responses.

However, fine-tuning also has its challenges. It requires careful management to avoid overfitting the training data and losing the ability to generalize to new situations. It also requires more computational resources and expertise to carry out effectively.

Both in-context learning and fine-tuning have their strengths and weaknesses. In-context learning is a cost-effective and straightforward approach that can be effective for simple tasks or when resources are limited. Fine-tuning, while more resource-intensive, can provide superior performance for complex tasks or when a high level of customization is desired.

The choice between the two will depend on your specific use case, the resources you have available, and the level of performance you require. It may also be beneficial to experiment with both approaches and evaluate their performance in your specific context.

Question 3

With image generative AI. Who owns the rights of the asset?

Accepted Answer

The intellectual property rights of assets generated by image generative AI are a complex and evolving issue. As generative AI becomes increasingly prevalent in creative industries, the legal implications remain unclear, especially in relation to copyright infringement, ownership of AI-generated works, and the use of unlicensed content in training data.

Generative AI, such as Stable Diffusion, Midjourney, or DALL·E 2, can produce remarkable visuals in various styles, and the quality and speed of creation often surpass average human performance. However, the legal framework for these AI-generated works is still being established.

Courts are currently grappling with how intellectual property laws should be applied to generative AI, and several cases have already been filed. Companies that use generative AI need to ensure they are in compliance with the law and take steps to mitigate potential risks. This includes using training data free from unlicensed content and developing ways to show the provenance of generated content.

The ownership of assets generated by image generative AI is not straightforward and is subject to ongoing legal interpretation and legislation. It’s crucial for companies to stay informed about the evolving legal landscape and take proactive steps to mitigate potential legal risks.

Question 4

Do you have a reference point of specificity of the factors you need to include with the prompts?

Accepted Answer

The specificity of prompts in AI models is contingent on several factors:

Task Complexity: For simple tasks, general prompts may suffice. However, complex tasks often require more specific prompts to guide the model toward the desired output.
Model Capabilities: Advanced models may require less explicit prompts as they can infer from context, while less sophisticated models may need more detailed instructions.
Desired Precision: If high precision is required in the output, the prompts should be more specific to reduce the scope for interpretation by the model.
Domain Specificity: Tasks in specialized domains may require more detailed prompts that include domain-specific language or concepts.
Model Customization: If the model has been fine-tuned on specific data, it may require less explicit prompts for tasks related to that data.

In practice, it’s often a process of trial and error, experimenting with different levels of prompt specificity to achieve optimal results.

Question 5

Isn't fine tuning only available for davinci, not gpt3.5/4?

Accepted Answer

Fine-tuning is currently only available for the base models of OpenAI’s GPT-3, including Davinci.

Question 6

How much are plus and other plans?  How do you find the best AI for different applications?

Accepted Answer

The cost of using the ChatGPT Plus version is $20 per month. This subscription provides general access to ChatGPT even during peak times, faster response times, access to the plugins and the browser version as well as to new features and improvements.

For the plugins used in the webinar, the primary one is Zapier. The pricing for Zapier varies depending on the plan you choose, and it’s best to refer to Zapier’s official pricing page for the most accurate and up-to-date information. The other plugins used in the webinar are free of charge.

If you’re interested in exploring more about AI applications, I recommend checking out our comprehensive article that covers over 90 different AI tools across various industries and fields. This resource can provide you with a broader understanding of how AI can be leveraged for different purposes and help you identify the best AI solutions for your specific needs.

Question 7

How to develop a custom plugin?

Accepted Answer

Creating a custom plugin for ChatGPT involves several steps, including building an API, creating a plugin manifest, and documenting the API. Here’s a more detailed process based on the guide from MLQ.ai:

Build an API: The first step is to create an API for your plugin. In the guide, they use the Quart framework to create a simple to-do list API. The API includes routes for creating, reading, and removing to-do lists for a user.
Create a Plugin Manifest: Every ChatGPT plugin needs a ai-plugin.json file that needs to be hosted on the API’s domain and located within a /.well-known folder. This file includes information about the plugin such as its name, description, authentication type, API type and URL, logo URL, contact email, and legal info URL.
Document the API with an OpenAPI Specification: You need to document your API with an OpenAPI specification. This specification includes information about the paths, operations, parameters, and responses of your API. The guide provides an example of an OpenAPI specification for the to-do list app.
Run the Plugin: After setting up the API, plugin manifest, and OpenAPI specification, you can run your plugin. In the guide, they run the plugin locally and then enable it in the ChatGPT plugin store by entering the local host URL.
Test the Plugin: Finally, you can test your plugin by enabling it in the installed plugins dropdown in ChatGPT and using it in a conversation.

Remember, this is a basic setup for testing purposes on localhost. For production apps, you’ll need to expose a public API for your app and set up user authentication.

Please note that this process might change as OpenAI continues to develop and refine the plugin system for ChatGPT. Always refer to the most recent documentation provided by OpenAI for the most accurate and up-to-date information.

How to Grow a Start-up

Daniil Stolyarov: Tech Titans & Investment

Top 8 Personal Injury Advertising Techniques That Really Work

How is Generative AI Shaping the Way We Work and Create

About the Generative AI Webinar

Transforming Your Work Process & Creativity with Generative AI

What You Will Learn

Webinar Agenda

Introduction of Generative AI and Its Possibilities

Optimizing AI Models and Prompts

Q&A Session

Meet our Speaker

How Will I Benefit?

Webinar Recording

Frequently Asked Questions