This article was accepted into the corpus but its outbound wikilinks were never NER-processed — typical at the deepest BFS hop or when the run's entity cap was reached. No expansion funnel to show.
| GRETAP | |
|---|---|
| Name | GRETAP |
| Type | Research framework |
| First released | 2023 |
| Developer | Consortium of research institutes |
| License | Open-source |
| Latest release | 2025 |
GRETAP GRETAP is a multimodal research platform and toolkit created to facilitate large-scale analysis, synthesis, and benchmarking of generative technologies. It integrates components for model orchestration, dataset curation, and evaluation across modalities with a focus on reproducibility and interoperability.
GRETAP provides a modular pipeline for coordinating datasets, model checkpoints, and evaluation suites developed by organizations such as OpenAI, DeepMind, Google Research, Meta AI, and Microsoft Research. It supports interoperability with standards and projects including Hugging Face, TensorFlow, PyTorch, OpenCV, and scikit-learn. Designed for academic and industrial laboratories like MIT CSAIL, Stanford AI Lab, Carnegie Mellon University, Berkeley AI Research, and ETH Zurich, GRETAP aims to bridge initiatives exemplified by ImageNet, COCO, LibriSpeech, Common Crawl, and C4. The project draws on practices from consortia such as Partnership on AI, IEEE Standards Association, and OpenAI Scholars to align with benchmarks used in GLUE, SuperGLUE, WMT, SQuAD, and BLEU.
Initial proposals for GRETAP emerged in meetings involving stakeholders from National Science Foundation, European Commission, DARPA, Alan Turing Institute, and private labs including Anthropic and Stability AI. Early development was influenced by reusable workflows from Kubernetes, Docker, Apache Airflow, and research reproducibility efforts at Yale University and Harvard University. Key milestones map to public releases and workshops at venues such as NeurIPS, ICML, ACL, CVPR, and ICLR. Governance adopted community models observed in Linux Foundation projects and guidelines discussed at UNESCO and OECD panels on artificial intelligence. Collaborations included dataset licensing negotiations referencing frameworks from Creative Commons, Open Data Institute, and legal experts from Harvard Law School and Stanford Law School.
GRETAP’s architecture composes microservices, registries, and evaluation primitives inspired by patterns in RESTful API deployments and orchestration used by Kubernetes and Istio. Core layers interface with model hosting systems such as TorchServe, TensorFlow Serving, ONNX Runtime, and hardware accelerators from NVIDIA, AMD, and Intel. Storage and retrieval systems integrate with distributed filesystems like Ceph and object stores modeled after Amazon S3 and Google Cloud Storage. Metadata, provenance, and lineage tracking follow practices from DataCite, PROV-O, and standards discussed at W3C workshops. The platform exposes SDKs and CLIs patterned on developer tools from GitHub, GitLab, and Bitbucket to enable contribution workflows like pull requests and continuous integration used in projects such as Travis CI and Jenkins.
GRETAP includes dataset ingestion pipelines compatible with formats employed by COCO, Open Images, ImageNet-22K, SQuAD, GLUE, and LibriVox. Model interoperability supports formats like ONNX, SavedModel, and serialized artifacts used by Hugging Face Transformers and fairseq. Evaluation modules implement metrics drawn from research including BLEU, ROUGE, METEOR, BERTScore, and domain-specific scorers used in competitions at Kaggle and ImageNet Large Scale Visual Recognition Challenge. Monitoring and visualization borrow interfaces from Prometheus, Grafana, TensorBoard, and deployment patterns seen in Seldon Core and KFServing. Collaboration features mirror practices from academic platforms such as Zenodo and Figshare.
Research groups at institutions like UC Berkeley, Caltech, University of Oxford, University of Cambridge, and Princeton University use GRETAP for reproducible experiments in multimodal learning, transfer learning, and robustness evaluation. Industry adopters including teams at Amazon, Apple, Meta Platforms, Google, and Microsoft apply it for benchmarking recommendation systems, vision-language models, and speech models aligned with datasets like MS COCO, VGGFace2, and Common Voice. GRETAP also serves cross-disciplinary projects spanning healthcare collaborations with Mayo Clinic, Johns Hopkins Medicine, and National Institutes of Health, environmental modeling with NASA and NOAA, and digital humanities initiatives at British Library and Bibliothèque nationale de France.
Deployment scenarios encompass on-premise clusters at research centers such as Lawrence Berkeley National Laboratory and cloud-hosted instances on platforms including Amazon Web Services, Google Cloud Platform, and Microsoft Azure. Continuous integration and reproducibility are supported through integrations with GitHub Actions, CircleCI, and research artifact registries similar to ArtifactHub. Hardware targets range from single-node GPU setups with NVIDIA A100 to multi-node TPU pods exemplified by deployments at Google TPU Research Cloud. Operators follow compliance and audit patterns used by institutions like ISO, NIST, and FedRAMP for regulated environments.
Security practices align with threat models and guidance from NIST Special Publication 800-53, OWASP, and privacy frameworks such as GDPR and HIPAA where applicable in collaborations with U.S. Department of Health and Human Services. Access controls leverage identity providers and protocols used by Okta, Auth0, OAuth 2.0, and SAML while encryption at rest and in transit follows standards from AES and TLS. Data governance workflows reflect policies discussed at Council of Europe and European Data Protection Board, and auditing trails implement provenance techniques referenced in work by DataCite and W3C PROV. Security-hardening draws on recommendations from vendors such as Cisco Security and CrowdStrike and incident response practices seen in CERT coordination centers.
Category:Research platforms