BEE-spoke-data/beecoder-playground

🐝BeeCoder Demo🐝

Code-Completion Playground 💻 with 🐝BeeCoder Models

This is a demo playground for generating Python code with the power of 🐝BeeCoder, a fine-tuned version of the tiny 101M base model on a dataset of pypi packages.

ℹ️ This is not an instruction model but just a code completion tool.

Intended Use: This app and its supporting model are provided for demonstration purposes only; not to serve as a replacement for human expertise. For more details on the model, please refer to the model card.

In our country, we say "To let 100M parameters model generate python script and not validate is like to let monkey fly a plane". So please be careful with the generated code.

Base Model Information

The base model, smol_llama-101M-GQA, has been pre-trained on a relatively small number of high quality tokens (less than ~20B). It has impressive performance despite its compact size of 101M parameters. Training data for this base model included:

• —JeanKaddour/minipile
• —pszemraj/simple_wikipedia_LM
• —BEE-spoke-data/wikipedia-20230901.en-deduped
• —mattymchen/refinedweb-3m

You can find more information about the base model here.

Credits

This app is modified from a demo playground originally built for StarCoder by BigCode. You can find the original demo here.