SantaCoder: don't reach for the stars!

TL;DR

This paper reports on the BigCode project's progress in developing large language models for code, highlighting data filtering techniques, model training on multiple languages, and outperforming larger models on code generation benchmarks.

Contribution

Introduces a 1.1B parameter multilingual code generation model trained on The Stack, with improved data filtering and evaluation showing superior performance over larger models.

Findings

Aggressive filtering of near-duplicate data improves performance.

Selecting files from repositories with fewer GitHub stars can enhance model quality.

The models outperform larger open-source models on code generation benchmarks.

Abstract

The BigCode project is an open-scientific collaboration working on the responsible development of large language models for code. This tech report describes the progress of the collaboration until December 2022, outlining the current state of the Personally Identifiable Information (PII) redaction pipeline, the experiments conducted to de-risk the model architecture, and the experiments investigating better preprocessing methods for the training data. We train 1.1B parameter models on the Java, JavaScript, and Python subsets of The Stack and evaluate them on the MultiPL-E text-to-code benchmark. We find that more aggressive filtering of near-duplicates can further boost performance and, surprisingly, that selecting files from repositories with 5+ GitHub stars deteriorates performance significantly. Our best model outperforms previous open-source multilingual code generation models (InCoder-6.7B and CodeGen-Multi-2.7B) in both left-to-right generation and infilling on the Java, JavaScript, and Python portions of MultiPL-E, despite being a substantially smaller model. All models are released under an OpenRAIL license at https://hf.co/bigcode.