Search for: All records

The BigCode community, an open-scientific collaboration working on the responsible development of Large Language Models for Code (Code LLMs), introduces StarCoder and StarCoderBase: 15.5B parameter models with 8K context length, infilling capabilities and fast large-batch inference enabled by multi-query attention. StarCoderBase is trained on 1 trillion tokens sourced from The Stack, a large collection of permissively licensed GitHub repositories with inspection tools and an opt-out process. We fine-tuned StarCoderBase on 35B Python tokens, resulting in the creation of StarCoder. We perform the most comprehensive evaluation of Code LLMs to date and show that StarCoderBase outperforms every open Code LLM that supports multiple programming languages and matches or outperforms the OpenAI code-cushman-001 model. Furthermore, StarCoder outperforms every model that is fine-tuned on Python, can be prompted to achieve 40% pass@1 on HumanEval, and still retains its performance on other programming languages. We take several important steps towards a safe open-access model release, including an improved PII redaction pipeline and a novel attribution tracing tool, and make the StarCoder models publicly available under a more commercially viable version of the Open Responsible AI Model license. 

Modern science generates large complicated heterogeneous collections of data. In order to effectively exploit these data, researchers must find relevant data, and enough of its associated metadata to understand it and put it in context. This problem exists across a wide range of research domains and is ripe for a general solution. Existing ventures address these issues using ad hoc purpose-built tools. These tools explicitly represent the data relationships by embedding them in their data storage mechanisms and in their applications. While producing useful tools, these approaches tend to be difficult to extend and data relationships are not necessarily traversable symmetrically. We are building a general system for navigational metadata. The relationships between data and between annotations and data are stored as first-class objects in the system. They can be viewed as instances drawn from a small set of graph types. General-purpose programs can be written which allow users explore these graphs and gain insights into their data. This process of data navigation, successive inclusion and filtering of objects provides powerful paradigm for data exploration.