Stable Code Technical Report

我们介绍了稳定代码（Stable Code），这是我们新一代代码语言模型系列中的第一个，它作为一个通用的基础代码语言模型，针对代码补全、推理、数学和其他软件工程任务。此外，我们还介绍了一种指令变体，名为稳定代码指令（Stable Code Instruct），它允许在自然对话界面中与模型进行交互，执行问答和基于指令的任务。在本技术报告中，我们详细介绍了导致这两个模型的数据和训练过程。它们的权重可以通过Hugging Face进行下载和使用，网址为https://huggingface.co/stabilityai/stable-code-3b和https://huggingface.co/stabilityai/stable-code-instruct-3b。本报告包含了对这些模型的彻底评估，包括多语言编程基准测试和以多轮对话为重点的MT基准测试。在发布时，稳定代码是3B参数下最先进的开放模型，甚至在流行的Multi-PL基准测试中表现与规模为70亿和150亿参数的更大模型相当。稳定代码指令在MT-Bench编码任务和Multi-PL补全等方面也表现出最先进的性能，与其他指令调整模型相比。鉴于其吸引人的小尺寸，我们还提供了多个边缘设备的吞吐量测量。此外，我们开源了几个量化检查点，并提供了它们与原始模型的性能指标比较。

Stable Code: First-Generation Open-Source Language Model for Code

The paper introduces Stable Code, a general-purpose base code language model and its instruction variant, Stable Code Instruct, for performing question-answering and instruction-based tasks. The models are trained on a large dataset and perform comparably to larger models on popular benchmarks.

The models are open-source and their weights are available for download. Thorough evaluations, including multilingual programming benchmarks and MT benchmark focusing on multi-turn dialogues, are presented. The models exhibit state-of-the-art performance and appealing small size, and throughput measurements on edge devices are provided. Quantized checkpoints are open-sourced and their performance metrics compared to the original model are provided.

Recent related work in this field includes GPT-3, Codex, and OpenAI's CodeBERT.