Workflow Diagram Explanation

This diagram illustrates the process of enhancing Large Language Models (LLMs) with custom adapters and integrating them into a user-friendly interface for both end-users and developers.

Alternative text for the image

The project’s pipeline.

Benchmarking

The workflow begins with the benchmarking phase, where the performance of various open-source LLMs is compared. The most suitable models are selected as baseline models for further development.

Alternative text for the image

Benchmarking various open source LLMs.

Fine-tuning & Adapter Creation

Next, the selected models undergo fine-tuning to optimize them for specific tasks. Custom adapters are then created to augment the LLMs’ abilities, which are saved on the Hugging Face platform, a hub for sharing machine learning models.

Hint

For more details, refer to the FineTuning ludwig tutorial.

Models with Adapters

These custom adapters are integrated into the baseline models. The integration allows for a more modular approach to machine learning model enhancement, where specific capabilities can be added without altering the entire model architecture.

Hint

For more details, refer to the Load adapter and attach to model tutorial.

Synthetic Data Generation

In parallel to adapter integration, synthetic data is generated using GPT-4. This data can be used to further train and refine the models, ensuring that they are well-equipped to handle a variety of scenarios.

Hint

For more details, refer to the Synthetic data tutorial.

Interface

Finally, the models with adapters are made accessible through two distinct interfaces:

  • User Interface Designed for end-users, this interface is user-friendly and allows users to select between default models or to perform custom fine-tuning with their own data.

  • Developer Interface Tailored for developers, this interface provides the tools needed to manage models, adapters, and test results effectively.

This diagram encapsulates the streamlined approach to adapting LLMs to specialized tasks and ensuring that both users and developers have the necessary tools at their disposal.

Note

In the final version of the interface, the Developer Interface will not be present, as the system is designed to be user-centric without requiring direct developer involvement.