Hypertuna

flowchart
step functions pipeline
the moment our hearts stopped :(

Inspiration

Current LLM customization workflows are slow and overly complex. Training an LLM for a task takes extremely long, and isn't viable for projects that need to be completed quickly (Like a hackathon!).

Additionally, the process of retrieving data, cleaning, setting up compute, training and hosting are tasks that require extensive knowledge of HuggingFace, AWS, and many other tools.

hypertuna is a hassle-free platform to simply ask for what sort of fine-tuning task is needed. Then, our data pipelines use RAG techniques to pull different data-sets, train, and then host your model to use in your apps.

What it does

Hypertuna is a streamlined LLM customization platform. Users provide a simple description of their task and select a base model. Hypertuna then auto-generates a fine-tuning job. Once fine-tuned, users can test the model directly in a web interface, or interact with the API

How we built it

We're using a wide variety of AWS suite tools to host our website!

We used DynamoDB for authentication and storing prompts for our model S3 was used for storage of datasets, and for the training data metrics Our biggest component: We used Step Functions to build the data pipeline that refines and fine-tunes a model based off a prompt. We used AWS Bedrock and Lambdas extensively here We used EC2 to communicate with all the components

Using step-functions helped us connect all of our data sources together,

Challenges we ran into

In the final stretch of day 2, we ran into a severe issue, our AWS accounts were not allowed to make Bedrock building jobs. :(

We had the entire system set up to be able to finish early, but we were unable to circumvent this roadblock without using tools like SageMaker, which did not fit our needs due to not being able to use arbitrary training data. This would also complicate the process of

This killed all of our plans, and we weren't expecting to run into this sort of roadblock. However, despite a blackout and stress from all-nighters. We pushed forward and pivoted our idea.

We set up the leading open-source models to work off of a bedrock-determined prompt, allowing our users to still get the benefits of a hassle-free model host. We managed to implement all our backend services in time to write this post ;)

What's next for hypertuna

The idea of supporting fine-tuning so quickly for developers is unheard of. As such, we plan to continue working on this project over the summer, consolidating our architecture and speeding up our pipeline to give developers an unheard-of product.