BERT-LLM Classification Fine-Tuning and Serving Pipeline

This repository contains a pipeline for fine-tuning a BERT-LLM model on a classification task and serving the model using the Union AI workflow and inference platform.

Project Setup

The quickest way to setup and the run the tutorial notebook is often using a hosted notebook, like Google colab.

Or you can follow the steps below to setup the project locally.

Sign up for a Union account

Serverless is the easiest way to get started with Union. You can sign up for a free account and $30 of credit at Union Serverless. BYOC (Bring Your Own Cloud) is also available for more features and advanced users. Schedule a demo to learn more about Union BYOC.

• Union Serverless Sign-up: https://www.union.ai/
• Union BYOC: https://docs.union.ai/byoc/user-guide/union-overview#union-byoc

Read more in the overview of Union Serverless and Union BYOC.

Clone the repository

git clone https://github.com/unionai-oss/bert-llm-classification-pipeline
cd bert-llm-classification-pipeline

Install dependencies

pip install -r requirements.txt

Authenticate to Union from CLI

After you have signed up for Union, you can authenticate to Union from the CLI.

If on Union Serverless union create login --serverless --auth device-flow

If on Union BYOC (Bring Your Own Cloud) union create login --host <union-host-url>

Now your environment is setup to run the project on remotely Union.

• Autneticatoin docs

Run through the tutorial notebook or run the pipeline from CLI.

The tutorial notebook will guide you through the steps to fine-tune a BERT-LLM model on a classification task and serve the model using Union AI.

jupyter notebook tutorial.ipynb

Or you can run the steps for the training pipeline and serving from the CLI.

Train the model:

Serve the model:

Run Batch Inference:

Run Near Real-time Batch Inference with Actors: