ajperry-pipeline 0.1.21

No project description provided

Kubeflow Pipelines

Purpose

The purpose of this repository is a place to learn new things. From infrastructure to new LLM training techniques.

State

Data: Currently I have a functional transformer pipeline which learns common n-grams in human speech. This is due to the lack of data primarily.

Model: The transformer model is tested, and has been compared to other implementation. It's prediction time is sensible and small enough for my local hardware.

Future Work

Improved ML: Currently only supervised learning is employed, I expect the performance will plateau without reinforcement learning with human feedback (RLHF). This is to be added to the reddit pipeline.

Improved Logging: After training I'd like a set of input output pairs logged to MlFlow for increased transparency to output deficiencies.

Pipelines

1. Reddit Iteratively learning to create engaging posts with reddit data.

Reddit

Baremetal Usage

1. Run notebook notebooks/reddit_training.ipynb
   • Define hyperparameters that make sense for your system
2. Metrics are recorded locally and can be observed with locally running mlflow or with the verbose=true options, test examples are printed to standard out

Kubeflow Usage

1. Upload notebook notebooks/reddit_pipeline.ipynb
2. Define environment variables
3. Run cells defining training pipeline
4. Run/Schedule pipeline

Pipeline Description

The pipeline is ran each day. In this process this is done:

• New data is downloaded
• The current best model is downloaded and evaluated
• If the model has degraded or is not proficient, training is ran

At the time of writing this I only have 500 samples in training set, so a test BLEU score of 0 is expected, though I hope in the coming days it will improve.

The pipeline records metrics in mlflow and records the hyperparameters/logs/outputs of each run.