The BigCode project is an open-scientific collaboration working on the responsible development of large language models for code. This tech report describes the progress of the collaboration until December 2022, outlining the current state of the Personally Identifiable Information (PII) redaction pipeline, the experiments conducted to de-risk the model architecture, and the experiments investigating better preprocessing methods for the training data. We train 1.1B parameter models on the Java, JavaScript, and Python subsets of The Stack and evaluate them on the MultiPL-E text-to-code benchmark. We find that more aggressive filtering of near-duplicates can further boost performance and, surprisingly, that selecting files from repositories with 5+ GitHub stars deteriorates performance significantly. Our best model outperforms previous open-source multilingual code generation models (InCoder-6.7B and CodeGen-Multi-2.7B) in both left-to-right generation and infilling on the Java, JavaScript, and Python portions of MultiPL-E, despite being a substantially smaller model. All models are released under an OpenRAIL license at https://hf.co/bigcode.

BACKGROUND: The overall survival of patients with advanced and refractory oesophageal squamous cell carcinoma, mostly aged 65 years and older, is poor. Treatment with PD-1 antibodies showed improved progression-free survival and overall survival. We assessed the safety and efficacy of combined nivolumab and ipilimumab therapy in this population. METHODS: This multicentre, open-label, phase 2 trial done in 32 sites in Germany included patients aged 65 years and older with oesophageal squamous cell carcinoma and disease progression or recurrence following first-line therapy. Patients were treated with nivolumab (240 mg fixed dose once every 2 weeks, intravenously) in the safety run-in phase and continued with nivolumab and ipilimumab (nivolumab 240 mg fixed dose once every 2 weeks and ipilimumab 1 mg/kg once every 6 weeks, intravenously). The primary endpoint was overall survival, which was compared with a historical cohort receiving standard chemotherapy in the intention-to-treat population. This study is registered with ClinicalTrials.gov, NCT03416244. FINDINGS: Between March 2, 2018, and Aug 20, 2020, we screened 75 patients with advanced oesophageal squamous cell carcinoma. We enrolled 66 patients (50 [76%] men and 16 [24%] women; median age 70·5 years [IQR 67·0-76·0]), 44 (67%) of whom received combined nivolumab and ipilimumab therapy and 22 (33%) received nivolumab alone. Median overall survival time at the prespecified data cutoff was 7·2 months (95% CI 5·7-12·4) and significantly higher than in a historical cohort receiving standard chemotherapy (p=0·0063). The most common treatment-related adverse events were fatigue (12 [29%] of 42), nausea (11 [26%]), and diarrhoea (ten [24%]). Grade 3-5 treatment-related adverse events occurred in 13 (20%) of 66 patients. Treatment-related death occurred in one patient with bronchiolitis obliterans while on nivolumab and ipilimumab treatment. INTERPRETATION: Patients aged at least 65 years, with advanced oesophageal squamous cell carcinoma might benefit from combined nivolumab and ipilimumab therapy in second-line treatment. FUNDING: Bristol Myers Squibb.

In this paper, we introduce AutoRDF2GML, a framework designed to convert RDF data into data representations tailored for graph machine learning tasks. AutoRDF2GML enables, for the first time, the creation of both content-based features---i.e., features based on RDF datatype properties---and topology-based features---i.e., features based on RDF object properties. Characterized by automated feature extraction, AutoRDF2GML makes it possible even for users less familiar with RDF and SPARQL to generate data representations ready for graph machine learning tasks, such as link prediction, node classification, and graph classification. Furthermore, we present four new benchmark datasets for graph machine learning, created from large RDF knowledge graphs using our framework. These datasets serve as valuable resources for evaluating graph machine learning approaches, such as graph neural networks. Overall, our framework effectively bridges the gap between the Graph Machine Learning and Semantic Web communities, paving the way for RDF-based machine learning applications.

The Webis MS MARCO Anchor Text 2022 dataset enriches Version 1 and 2 of the document collection of MS MARCO with anchor text extracted from six Common Crawl snapshots. The six Common Crawl snapshots cover the years 2016 to 2021 (between 1.7-3.4 billion documents each). We sampled 1,000 anchor texts for documents with more than 1,000 anchor texts at random and all anchor texts for documents with less than 1,000 anchor texts (this sampling yields that all anchor text is included for 94% of the documents in Version 1 and 97% of documents for Version 2). Overall, the MS MARCO Anchor Text 2022 dataset enriches 1,703,834 documents for Version 1 and 4,821,244 documents for Version 2 with anchor text. Cleaned versions of the MS MARCO Anchor Text 2022 dataset are available in ir_datasets, Zenodo and Hugging Face. The raw dataset with additional information and all metadata for the extracted anchor texts (roughly 100GB) is available on Hugging Face and files.webis.de. The details of the construction of the Webis MS MARCO Anchor Text 2022 dataset are described in the associated paper. If you use this dataset, please cite @InProceedingsfroebe:2022a, address = Berlin Heidelberg New York, author = Maik Fr``obe and Sebastian G''unther and Maximilian Probst and Martin Potthast and Matthias Hagen, booktitle = Advances in Information Retrieval. 44th European Conference on IR Research (ECIR 2022), editor = Matthias Hagen and Suzan Verberne and Craig Macdonald and Christin Seifert and Krisztian Balog and Kjetil Norvrag and Vinay Setty, month = apr, publisher = Springer, series = Lecture Notes in Computer Science, site = Stavanger, Norway, title = The Power of Anchor Text in the Neural Retrieval Era, year = 2022

AIMS: Artificial intelligence (AI) provides a powerful tool to extract information from digitised histopathology whole slide images. During the last 5 years, academic and commercial actors have developed new technical solutions for a diverse set of tasks, including tissue segmentation, cell detection, mutation prediction, prognostication and prediction of treatment response. In the light of limited overall resources, it is presently unclear for researchers, practitioners and policymakers which of these topics are stable enough for clinical use in the near future and which topics are still experimental, but worth investing time and effort into. METHODS AND RESULTS: To identify potentially promising applications of AI in pathology, we performed an anonymous online survey of 75 computational pathology domain experts from academia and industry. Participants enrolled in 2021 were queried about their subjective opinion on promising and appealing subfields of computational pathology with a focus upon solid tumours. The results of this survey indicate that the prediction of treatment response directly from routine pathology slides is regarded as the most promising future application. This item was ranked highest in the overall analysis and in subgroups by age and professional background. Furthermore, prediction of genetic alterations, gene expression and survival directly from routine pathology images scored consistently high throughout subgroups. CONCLUSIONS: Together, these data demonstrate a possible direction for the development of computational pathology systems in clinical, academic and industrial research in the near future.

The ImageCLEF evaluation campaign was integrated with CLEF (Conference and Labs of the Evaluation Forum) for more than 20 years and represents a Multimedia Retrieval challenge aimed at evaluating the technologies for annotation, indexing, and retrieval of multimodal data. Thus, it provides information access to large data collections in usage scenarios and domains such as medicine, argumentation and content recommendation. ImageCLEF 2024 has four main tasks: (i) a Medical task targeting automatic image captioning for radiology images, synthetic medical images created with Generative Adversarial Networks (GANs), Visual Question Answering and medical image generation based on text input, and multimodal dermatology response generation; (ii) a joint ImageCLEF-Touché task Image Retrieval/Generation for Arguments to convey the premise of an argument, (iii) a Recommending task addressing cultural heritage content-recommendation, and (iv) a joint ImageCLEF-ToPicto task aiming to provide a translation in pictograms from natural language. In 2023, participation increased by 67% with respect to 2022 which reveals its impact on the community.

With climate extremes' rising frequency and intensity, robust analytical tools are crucial to predict their impacts on terrestrial ecosystems. Machine learning techniques show promise but require well-structured, high-quality, and curated analysis-ready datasets. Earth observation datasets comprehensively monitor ecosystem dynamics and responses to climatic extremes, yet the data complexity can challenge the effectiveness of machine learning models. Despite recent progress in deep learning to ecosystem monitoring, there is a need for datasets specifically designed to analyse compound heatwave and drought extreme impact. Here, we introduce the DeepExtremeCubes database, tailored to map around these extremes, focusing on persistent natural vegetation. It comprises over 40,000 globally sampled small data cubes (i.e. minicubes), with a spatial coverage of 2.5 by 2.5 km. Each minicube includes (i) Sentinel-2 L2A images, (ii) ERA5-Land variables and generated extreme event cube covering 2016 to 2022, and (iii) ancillary land cover and topography maps. The paper aims to (1) streamline data accessibility, structuring, pre-processing, and enhance scientific reproducibility, and (2) facilitate biosphere dynamics forecasting in response to compound extremes.

Mesenchymal stem cell therapy shows promising results for difficult-to-treat diseases, but standardized manufacturing requires robust quality control through automated cell confluence monitoring. While deep learning can automate confluence estimation, research on cost-effective dataset curation and the role of foundation models in this task remains limited. We systematically investigate the most effective strategies for confluence estimation, focusing on active learning-based dataset curation, goal-specific labeling, and leveraging foundation models for zero-shot inference. Here, we show that zero-shot inference with the Segment Anything Model (SAM) achieves excellent confluence estimation without any task-specific training, outperforming fine-tuned smaller models. Further, our findings demonstrate that active learning does not significantly improve model dataset curation compared to random selection in homogeneous cell datasets. We show that goal-specific, simplified labeling strategies perform comparably to precise annotations while substantially reducing annotation effort. These results challenge common assumptions about dataset curation: neither active learning nor extensive fine-tuning provided significant benefits for our specific use case. Instead, we found that leveraging SAMtextquoterights zero-shot capabilities and targeted labeling strategies offers the most cost-effective approach to automated confluence estimation. Our work provides practical guidelines for implementing automated cell monitoring in MSC manufacturing, demonstrating that extensive dataset curation may be unnecessary when foundation models can effectively handle the task out of the box.Competing Interest StatementThe authors have declared no competing interest.

textcopyright 2021 Copyright for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).Idiosyncrasies in human writing styles make it difficult to develop systems for authorship identification that scale well across individuals. In this year's edition of PAN, the authorship identification track focused on open-set authorship verification, so that systems are applied to unknown documents by previously unseen authors in a new domain. As in the previous year, the sizable materials for this campaign were sampled from English-language fanfiction. The calibration materials handed out to the participants were the same as last year, but a new test set was compiled with authors and fandom domains not present in any of the previous datasets. The general setup of the task did not change, i.e., systems still had to estimate the probability of a pair of documents being authored by the same person. We attracted 13 submissions by 10 international teams, which were compared to three complementary baselines, using five diverse evaluation metrics. Post-hoc analyses show that systems benefitted from the abundant calibration materials and were well-equipped to handle the open-set scenario: Both the top-performing approach and the highly competitive cohort of runner-ups presented surprisingly strong verifiers. We conclude that, at least within this specific text variety, (large-scale) open-set authorship verification is not necessarily or inherently more difficult than a closed-set setup, which offers encouraging perspectives for the future of the field.

Each year the International Semantic Web Conference accepts a set of Semantic Web Challenges to establish competitions that will advance the state of the art solutions in any given problem domain. The SeMantic AnsweR Type prediction task (SMART) was part of ISWC 2020 challenges. Question type and answer type prediction can play a key role in knowledge base question answering systems providing insights that are helpful to generate correct queries or rank the answer candidates. More concretely, given a question in natural language, the task of SMART challenge is, to predict the answer type using a target ontology (e.g., DBpedia or Wikidata).

AbstractTemporal property graphs are graphs whose structure and properties change over time. Temporal graph datasets tend to be large due to stored historical information, asking for scalable analysis capabilities. We give a complete overview of Gradoop, a graph dataflow system for scalable, distributed analytics of temporal property graphs which has been continuously developed since 2005. Its graph model TPGM allows bitemporal modeling not only of vertices and edges but also of graph collections. A declarative analytical language called GrALa allows analysts to flexibly define analytical graph workflows by composing different operators that support temporal graph analysis. Built on a distributed dataflow system, large temporal graphs can be processed on a shared-nothing cluster. We present the system architecture of Gradoop, its data model TPGM with composable temporal graph operators, like snapshot, difference, pattern matching, graph grouping and several implementation details. We evaluate the performance and scalability of selected operators and a composed workflow for synthetic and real-world temporal graphs with up to 283 M vertices and 1.8 B edges, and a graph lifetime of about 8 years with up to 20 M new edges per year. We also reflect on lessons learned from the Gradoop effort.

Machine learning models for multi-omics data often trade off predictive accuracy against biological interpretability. An emerging class of deep learning architectures structurally encode biological knowledge to improve both prediction and explainability. Opportunities and challenges remain for broader adoption.

Immunotherapy is currently one of the most promising treatment options for malignant melanoma. To uncover new immunological targets for future treatment approaches, single-cell transcriptomic and epigenomic analyses were performed on human primary melanoma and melanocytic nevus samples. The detailed methods of this study are described in the Supplementary Material.

Abstract argumentation frameworks are by now a major research area in knowledge representation and reasoning. Various aspects of AFs have been extensively studied over the last 25 years. Contributing to understanding the expressive power of AFs, researchers found lower and upper bounds for the maximal number of extensions, that is, acceptable points of view, in AFs. One of the classical and most important concepts in AFs are so-called complete extensions. Surprisingly, the exact bound for the maximal number of complete extensions in an AF has not yet been formally established, although there is a reasonable conjecture tracing back at least to 2015. Recently the notion of modularization was introduced and it was shown that this concept plays a key role for the understanding of relations between semantics as well as intrinsic properties. In this paper, we will use this property to give a formal proof of the conjecture regarding complete semantics.

Monitoring the status of large computing systems is essential to identify unexpected behavior and improve their performance and uptime. However, due to the large-scale and distributed design of such computing systems as well as a large number of monitoring parameters, automated monitoring methods should be applied. Such automatic monitoring methods should also have the ability to adapt themselves to the continuous changes in the computing system. In addition, they should be able to identify behavioral anomalies in useful time, to perform appropriate reactions. This work proposes a general lightweight and unsupervised method for near real-time anomaly detection using operational data measurement on large computing systems. The proposed model requires as little as 4 hours of data and 50 epochs for each training process to accurately resemble the behavioral pattern of computing systems.

Lipocalin family proteins have been shown to bind a vast array of small molecules and have subsequently been adapted to selectively bind specific ligands. In this study, candesartan, an antihypertension drug, was identified to bind mouse and human siderocalin in biomolecular NMR experiments, allowing for structural insights into the candesartan-siderocalin interaction. The ligand binding site was determined through an integrative structural biology approach using in silico ligand docking guided by NMR experiments. Building on this structurally informed binding model, we used rational protein design to modulate the binding pocket for increased or decreased ligand binding affinity. The predicted mutations were evaluated in vitro using isothermal titration calorimetry. This resulted in a mutant with a 50-fold increase in binding affinity in addition to a second mutant with a five-fold decrease in binding affinity. Thus, siderocalins have potential as a scaffold for creation of various ligand binding-based tools, including drug scavengers.

The advent of endovascular thrombectomy has significantly improved outcomes for stroke patients with intracranial large vessel occlusion, yet individual benefits can vary widely. As demand for thrombectomy rises and geographic disparities in stroke care access persist, there is a growing need for predictive models that quantify individual benefits. However, current imaging methods for estimating outcomes may not fully capture the dynamic nature of cerebral ischemia and lack a patient-specific assessment of thrombectomy benefits. Our study introduces a deep learning approach to predict individual responses to thrombectomy in acute ischemic stroke patients. The proposed models provide predictions for both tissue and clinical outcomes under two scenarios: one assuming successful reperfusion and another assuming unsuccessful reperfusion. The resulting simulations of penumbral salvage and difference in NIHSS at discharge quantify the potential individual benefits of the intervention. Our models were developed on an extensive dataset from routine stroke care, which included 405 ischemic stroke patients who underwent thrombectomy. We used acute data for training (n = 304), including multimodal CT imaging and clinical characteristics, along with post hoc markers like thrombectomy success, final infarct localization, and NIHSS at discharge. We benchmarked our tissue outcome predictions under the observed reperfusion scenario against a thresholding-based clinical method and a generalised linear model. Our deep-learning model showed significant superiority, with a mean Dice score of 0.48 on internal (n = 50) and 0.52 on external (n = 51) test data, versus 0.26/0.36 and 0.34/0.35 for the baselines, respectively. The NIHSS sum score prediction achieved median absolute errors of 1.5 NIHSS points on the internal test dataset and 3.0 NIHSS points on the external test dataset, outperforming other machine learning models. By predicting the patient-specific response to thrombectomy for both tissue and clinical outcomes, our approach offers an innovative biomarker that captures the dynamics of cerebral ischemia. We believe this method holds significant potential to enhance personalised therapeutic strategies and to facilitate efficient resource allocation in acute stroke care.

Wikidata, now a decade old, is the largest public knowledge graph, with data on more than 100 million concepts contributed by over 560,000 editors. It is widely used in applications and research. At its launch in late 2012, however, it was little more than a hopeful new Wikimedia project, with no content, almost no community, and a severely restricted platform. Seven years earlier still, in 2005, it was merely a rough idea of a few PhD students, a conceptual nucleus that had yet to pick up many important influences from others to turn into what is now called Wikidata. In this paper, we try to recount this remarkable journey, and we review what has been accomplished, what has been given up on, and what is yet left to do for the future.

Problem-solving strategies have been investigated in various informatics education contexts. However, no substantial research has yet been conducted on the problem-solving behavior of students in the field of Machine Learning (ML). This study aims to bridge this gap by analyzing the self-directed problem-solving processes employed by students in grades 8 to 10 $$(n = 93)$$(n=93)when developing decision trees as classification models. A digital multi-touch puzzle game and a custom-developed toolchain were utilized to visually capture and subsequently analyze students' gameplay behaviors using quantitative content analysis techniques. The results of this study indicate that learners within the examined age group predominantly employ exploratory problem-solving strategies in the self-directed construction of decision trees. In contrast, structured approaches are employed much less frequently and demonstrate lower persistence, yet they are significantly more correlated with successful game completion. These findings underscore the necessity of developing learning environments that promote the application and facilitate the persistence of structured problem-solving strategies, enabling learners to engage with the functioning and development of decision trees in a systematic and purposeful manner.