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. 

Deep neural networks implementing generative models for dimensionality reduction have been extensively used for the visualization and analysis of genomic data. One of their key limitations is lack of interpretability: it is challenging to quantitatively identify which input features are used to construct the embedding dimensions, thus preventing insight into why cells are organized in a particular data visualization, for example. Here we present a scalable, interpretable variational autoencoder (siVAE) that is interpretable by design: it learns feature embeddings that guide the interpretation of the cell embeddings in a manner analogous to factor loadings of factor analysis. siVAE is as powerful and nearly as fast to train as the standard VAE but achieves full interpretability of the embedding dimensions. Using siVAE, we exploit a number of connections between dimensionality reduction and gene network inference to identify gene neighborhoods and gene hubs, without the explicit need for gene network inference. We observe a systematic difference in the gene neighborhoods identified by dimensionality reduction methods and gene network inference algorithms in general, suggesting they provide complementary information about the underlying structure of the gene co-expression network. Finally, we apply siVAE to implicitly learn gene networks for individual iPSC lines and uncover a correlation between neuronal differentiation efficiency and loss of co-expression of several mitochondrial complexes, including NADH dehydrogenase, cytochrome C oxidase, and cytochrome b.