Microsoft Researchers Introduce CodeOcean and WaveCode

Recent advances in AI, particularly in the area of ‚Äč‚Äčlarge language models (LLM), have led to remarkable developments in code language models. Microsoft researchers have introduced two innovative tools in this field: WaveCoder and CodeOcean, marking a significant leap forward in the field of instruction setting for code language models.

WaveCoder: Fine-tuned code LLM

WaveCoder is a fine-tuned Code Language Model (Code LLM) designed specifically to improve instruction tuning. The model demonstrates superior performance on a variety of code-related tasks, consistently outperforming other open-source models at the same level of fine-tuning. WaveCoder’s performance is particularly noticeable for tasks such as code generation, repair, and summarization.

CodeOcean: A rich dataset for improved instruction tuning

CodeOcean, the centerpiece of this research, is a carefully curated dataset comprising 20,000 instruction instances in four critical code-related tasks: code summarization, code generation, code translation, and code repair. Its main purpose is to increase the performance of Code LLM by fine-tuning instructions. CodeOcean excels by focusing on data quality and diversity, delivering superior performance across a variety of code-related tasks.

A new approach to instruction setting

The innovation lies in the method of using diverse, high-quality instruction data from open source code to revolutionize instruction tuning. This approach addresses the challenges of generating instructional data, such as the presence of duplicate data and limited data quality control. By categorizing the instruction data into four universal code-related tasks and refining the instruction data, the researchers have created a robust method to improve the generalizability of fine-tuned models.

The importance of data quality and diversity

This groundbreaking research highlights the importance of data quality and diversity in instructional setting. The new LLM-based Generator-Discriminator framework uses source code, allowing clear control over data quality during the generation process. This methodology excels at generating more authentic instruction data, thereby improving the generalizability of fine-tuned models.

WaveCoder benchmark performance

WaveCoder models have been rigorously evaluated in various fields, confirming their efficacy in various scenarios. They consistently outperform their peers in multiple benchmarks, including HumanEval, MBPP, and HumanEvalPack. A comparison with the CodeAlpaca dataset highlights CodeOcean’s superiority in refining the instruction data and increasing the instruction-following insight of the underlying models.

Implications for the market

For the market, Microsoft’s CodeOcean and WaveCoder herald a new era of more capable and adaptable coding language models. These innovations offer improved solutions for a range of applications and industries, enhancing the generalizability of LLMs and expanding their applicability across contexts.

Future directions

Further improvements in the monotask performance and generalization ability of the model are expected in the future. Interaction between different tasks and larger data sets will be key areas of focus to continue to advance the field of instruction tuning for code language models.


Microsoft’s introduction of WaveCoder and CodeOcean represents a major moment in the evolution of code language models. By emphasizing data quality and instruction set diversity, these tools pave the way for more complex, efficient, and adaptive models that are better equipped to handle a wide range of code-related tasks. This research not only improves the capabilities of large language models, but also opens new avenues for their application in various industries, marking an important milestone in the field of artificial intelligence.

Image source: Shutterstock

Leave a Comment