IBM Open Sources Granite Code Model

TapTechNews May 21 news, IBM announced at the annual THINK conference that IBM has now open-sourced its "most advanced" and high-performance Granite code model.

TapTechNews attached the open source links as follows:

GitHub: Click here to enter

HuggingFace: Click here to enter

The Granite code model has a parameter range from 3B to 34B, and there are two variants of the basic model and the instruction-following model, which are suitable for tasks such as complex application modernization, code generation, error repair, interpretation and documentation of code, and maintenance of databases. These code models are trained on 116 programming languages, and always reach the "most advanced" performance of the open source code large model in various code-related tasks.

IBM said that the Granite code model performs extremely well overall in all model sizes and benchmarks, usually outperforming other open source code models, even though the scale of these models is twice that of Granite. The Granite code model performs well in code synthesis, repair, interpretation, editing, and translation, covering most major programming languages including Python, JavaScript, Java, Go, C++, and Rust.

In addition, the 20B parameter Granite basic code model is used to train IBMwatsonxCodeAssistant (WCA) to adapt to professional fields. The 20B parameter Granite basic code model is tuned to generate SQL from natural language questions in order to transform structured data and extract insights (extracting insights).