December 2024

The Nova models are available through the BedRock API.

All three models can ingest and output text and code. Lite and Pro can, also, ingest images, video and various document formats.

Pro and Lite have an input context window of 300k tokens; Lite has ones of 128k tokens. All 3 models have a max output of 5k tokens. Additional guidance is outlines in the User Guide.

Proprietary. Users must follow AWS [Terms of Service](https://aws.amazon.com/service-terms/).

The User Guide has detailed instructions for use with examples.

The Nova models are clearly and accessibly documented: The User Guide provides detailed guidance and examples for each of the Nova models. The report contains detailed explanations of the range of evaluations conducted.

Guidelines for Responsible Use: https://docs.aws.amazon.com/nova/latest/userguide/responsible-use.html

AWS does not use inputs and outputs generated from NOVA models for training.

Data takedown requests can be submitted here: https://titan.aws.com/privacy

The Nova models are developed and tested in accordance with the following Responsible AI Dimensions: Fairness, Explainability, Privacy and Security, Safety, Controllability, Veracity and robustness, Governance and Transparency.

Issues can be reported inside of Bedrock in the AWS console.

The Nova models can be used for image, video, document and text understanding and text/code generation. They, also, support [tool use](https://docs.aws.amazon.com/nova/latest/userguide/tool-use.html) for calling specific APIs or code functions. Key limitations are outlined in the User Guide. For example, the Nova models do not support people identification in images and may struggle with multi-lingual image understanding; in video inputs, audio and timestamps are not supported.

Nova Pro, Lite and Micro are Transformed-based models.

The models were pretrained using a mixture of multilingual and multimodal data. Next, the models were post-trained using Supervised Fine-Tuning and reward model training on human preference data. Finally, methods like Direct Preference Optimization and Proximal Policy Optimization were used to ensure alignment with human preference.

Input and output moderation models are used to detect unsafe prompts and generated content.

The Nova family of models were trained on Amazon’s custom Trainium1 (TRN1) chips,10 NVidia A100 (P4d instances), and H100 (P5 instances) accelerators.

No information provided.

The Nova models were trained on a variety of sources, including licensed data, proprietary data, open source datasets, and publicly available data. This included text data for over 200 languages including Arabic, Dutch, English, French, German, Hebrew, Hindi, Italian, Japanese, Korean, Portuguese, Russian, Simplified Chinese, Spanish, and Turkish. For post-training, new data were created to demonstrated safe behavior based on the Amazon RAI objectives.

The curation of RLHF datasets is discussed, but no details on labor used were identified.

The pre-training data was curated for alignment with RAI objectives. This included de-identifying or removing certain types of personal data.

Certain types of personal data were removed or de-identified.

The Nova Technical Report includes the results from multiple benchmarks that cover multiple core capabilities, including text-only and multi-modal reasoning, agentic workflows and long-context text retrieval. While the exact evaluation code is not published, the prompts used included in the appendix. Qualitative examples of model performance on a multi-modal are included.

The Nova models are evaluated for Hallucinations and for alignment with RAI objectives. The Report explains the process for measuring alignment with RAI objectives, but does not include specific details.

The model developers framed the Adversarial Testing process around their 8 core Responsible AI dimensions: Fairness, Explainability, Privacy and Security, Safety, Controllability, Veracity and Robustness, Governance, and Transparency. Nova’s adherence to these principles was evaluated using automated benchmarks (public and new proprietary ones) and red-teaming (internal and external). The internal red-teaming exercise used a team of data analysts and subject-matter experts who tested the model’s robustness against adversarial prompts across all the RAI dimensions. The process resulted in a collection of over 300 adversarial prompting techniques that covered different languages and modalities. In addition, four external red-teaming groups (ActiveFence, Deloitte, Gomes Group and Nemysis) were employed to test the model in areas including hate speech, political misinformation and CBRN capabilities.

The developers used Supervised Fine Tuning (SFT) and Reinforcement Learning from Human Feedback (RLHF) methods to align the NOVA models with Amazon's RAI dimensions. The post-training datasets included single and multi-turn examples of expected safe behavior for each RAI dimension in multiple languages. In addition to alignment, the deployed Nova systems uses input and output moderation models that serve as an additional defense and allow the developers to respond more quickly to new threats.