Transforms

Overview

Entry Point's transform feature allows you to generate new fields in your dataset by running a prompt en masse against an existing field. For example, the real-world training data you have access to may consist of hundreds of example user inputs but no corresponding outputs, or vice versa.

Remember that the process of fine-tuning an LLM distills down to showing the model pairs of inputs and outputs; in a case where you only have one or the other, transforms allow you to synthesize the missing or complementary data quickly and easily.

Let's look at a couple of different use cases to help illustrate this.

Use Cases

Generating Inputs

First, consider a situation where your dataset consists of outputs only. For example, you may have dozens, hundreds, or even thousands of example social media posts representing the style, tone, and format of those that you'd like your fine-tuned model to create.

While having these example outputs is a good start, a model is not typically fine-tuned on outputs alone. Instead, each example output needs to be paired with an example user input for which you would expect the desired output to be produced by the model.

The table below illustrates a dataset containing outputs only. Transforms allow you to easily create the missing inputs that you need to complete your dataset.

Input (User Prompt)	Output (Social media post)
__	Leadership means embracing vulnerability and fostering a safe space for ideas and concerns. Show you're human, build trust, and drive innovation. #Leadership #Vulnerability #TrustBuilding
__	Active listening is key. Give your team full attention, acknowledge their views, and respond thoughtfully. This boosts morale and fosters mutual respect. #Leadership #ActiveListening #TeamEngagement
__	Lead by example. Your actions set the tone for your team. Show integrity, accountability, and dedication to inspire the same standards in others. #Leadership #LeadByExample #Integrity

Generating Outputs

Conversely, your training data may consist entirely of inputs for which you need to generate example outputs. For example, let's say your goal is to fine-tune a model to categorize inbound support requests. You may have access to plenty of example customer inquiries but not their corresponding categorizations.

Input (Customer Inquiry)	Output (Categorization)
I received the wrong item in my order, can you help?	__
My package says delivered, but I haven't received it.	__
The item I ordered arrived damaged, what are my options?	__

Whether your training dataset is missing inputs or outputs, Entry Points transform function can help. Instead of manually filling in the missing data, transforms allow you to leverage the power of LLMs to quickly and easily synthesize the data you need.

Other Use Cases

Transforms can also be used to generate meta data about your data. For instance, you could use a transform to populate a field that tells you if your inputs and outputs have issues.

See the Transform Data section of our Build a Dataset guide for more information on how to use transforms.