rotationalio/huggingface-example

Sample NLP streaming workflow using an LLM from HuggingFace and PyEnsign

huggingface-example

Sample NLP streaming workflow using an LLM from Hugging Face and PyEnsign

This is an example of a sentiment analysis application using sample yelp ratings data from Kaggle using Hugging Face and PyEnsign.

To use PyEnsign, create a free account on rotational.app. You will need to do the following once you create an account:

• Create a project.
• Add the following topic to the project: yelp_data. Check out this video on how to add a topic. You can choose your own names for the topic but make sure that you update the code accordingly.
• Generate API keys for your project.

You will need to create and source the following environment variables prior to running the example:

export ENSIGN_CLIENT_ID="your client id here"
export ENSIGN_CLIENT_SECRET="your client secret here"

This application consists of three components:

• Trainer reads data from the yelp_train.csv file and builds a model using the pretrained DistilBERT LLM from Hugging Face. The best model gets written to the final_model directory.
• ScoreDataPublisher reads data from the yelp_score.csv file publishes to the yelp_data topic.
• Scorer listens for new messages in the yelp_data topic. When it receives a new message, it uses the trained Hugging Face model in the final_model directory to make predictions.

Steps to run the application

Create a virtual environment

virtualenv venv

Activate the virtual environment

source venv/bin/activate

Install the required packages

pip install -r requirements.txt

Open three terminal windows

Run the Trainer in the first window (make sure to activate the virtual environment first). This will create three checkpoint directories under the trained_models directory and the final model configurations and weights in the final_model directory.

source venv/bin/activate

python huggingface_trainer.py

Once the training is complete, run the Scorer in the second window (make sure to activate the virtual environment first)

$ source venv/bin/activate

python huggingface_scorer.py score

Run the ScoreDataPublisher in the third window (make sure to activate the virtual environment first)

source venv/bin/activate

python huggingface_scorer.py score_data