Success with Fine-Tuning

For those in a hurry

AI models like OpenAI's GPT can be optimized with corporate data using fine-tuning. This increases the recognition rate for difficult cases to over 90%, sometimes even over 95% — excellent results.
During fine-tuning, a pre-trained AI model is trained with specific data to make it more sensitive to concrete tasks. It is therefore trained using data from specific and relevant business transactions.
In the case of GPT, fine-tuning is performed using prepared prompts that include the desired answers to the questions asked. The dataset containing corporate data to be analyzed is also provided so that the prompt can be applied to it.
Fine-tuning does not replace a clean data foundation or prior optimization of queries to achieve an excellent baseline quality. Fine-tuning is particularly suitable for optimizing an LLM model.

Tip for trying it out

If you want to create professional, multilingual explanatory or sales videos, you should take a closer look at the two platforms synthesia.io and HeyGen. While synthesia.io offers numerous different human avatars that present the desired texts with or without PowerPoint overlays, HeyGen is used for individual, multilingual videos based on your own persona. Both solutions are ideal for professional videos and demonstrate a high level of innovation.

Fine-Tuning - Advanced AI Training

Today, we find two types of AI models: pre-trained and non-pre-trained. While pre-trained models—the most prominent example being OpenAI GPT—have been trained using an immeasurable amount of text and multimedia content from the internet, non-pre-trained or "empty" models rely on professional users to calibrate or train them on the type of data for which they intend to use the model.

However, there is a middle ground that combines the advantages of both approaches: Fine-Tuning.

Fine-tuning builds upon a pre-trained model like OpenAI GPT and optimizes it with data relevant to the specific use case. This data can include documents (PDF, Word, Excel from SharePoint or cloud storage like Box.com, etc.), as well as data from an ERP system such as SAP, NetSuite, Xentral, MS Dynamics, or any other proprietary corporate database.

The Fine-Tuning Method Explained

Fine-tuning relies on the same mechanisms used in prompting for Large Language Models (LLMs). This means a question or instruction (prompt) is formulated, and the expected answer (completion) is provided. We are therefore dealing with prompt-completion pairs that are loaded into the GPT model via an interface.

However, since the AI model is to be optimized for specific data rather than the broad data foundation used for the base training, the underlying (corporate) data must also be provided alongside these prompt-completion pairs. Consequently, fine-tuning is carried out using many prompt-completion pairs and the corresponding dataset for each pair.

Once training is complete, the fine-tuned AI model is available to the user. In the case of OpenAI, a new API is generated that can be used both programmatically and via the visual ChatGPT frontend. Technically speaking, we have created a way to use our own model parameterization for our specific use case alongside the AI model's algorithm.

The Success of Optimization

Fine-tuning is particularly effective when well-developed prompts and perfectly configured AI platforms need to be extended to include business-specific use cases. In one of our projects, we used order data of a specific content type for fine-tuning, where the information was completely unstructured and scattered across PDF documents.

Through fine-tuning, we informed the AI model which answers we would expect based on the data and integrated this data into the fine-tuning process.

Results:

In the correct classification of line item descriptions, we increased the baseline quality of these content-heterogeneous fields from a 63% recognition rate to 91% — and that with only 4 fine-tuning datasets!
In the recognition of multi-stop addresses during delivery, the 78% recognition rate was increased to 96%.

In both cases, OpenAI's GPT-3.5 Turbo was used. We expect the recognition rates of GPT-4 Turbo to be even higher.

Is Fine-Tuning a Magic Bullet?

No! Even with fine-tuning, the prompts, the datasets selected for fine-tuning, the model selection, and supplementary measures such as Interactive Contextual Prompting ("few-shots", "one-shots") must be carefully considered and precisely executed. If the baseline quality of the model operation is not right, any fine-tuning is a gamble.

Supplementary Techniques:

Retrieval Augmented Generation (RAG) is generally used to complement fine-tuning. It allows large amounts of corporate data to be dynamically provided to a GPT model for processing, utilizing vector databases and embeddings.
Result validation using discrete methods (e.g., RegEx) or lookups (e.g., plausibility checks), or by employing additional, specialized AI models.

When well-thought-out, professionally tested, and carefully implemented, fine-tuned AI models can be perfectly utilized for automating complex business processes—from order entry and order planning to billing.