Events & Conferences
Auto-translating “Dive into Deep Learning” with Amazon Translate
Dive into Deep Learning (D2L.ai) is an open-source textbook that makes deep learning accessible to everyone. It features interactive Jupyter notebooks with self-contained code in PyTorch, JAX, TensorFlow, and MXNet, as well as real-world examples, exposition figures, and math. So far, D2L has been adopted by more than 400 universities around the world, such as the University of Cambridge, Stanford University, the Massachusetts Institute of Technology, Carnegie Mellon University, and Tsinghua University.
The latest updates to “Dive into Deep Learning”
Learn about the newest additions to the popular open-source, interactive book, including the addition of a Google JAX implementation and three new chapters in volume 2.
As a result of the book’s widespread adoption, a community of contributors has formed to work on translations in various languages, including Chinese, Japanese, Korean, Portuguese, Turkish, and Vietnamese. To efficiently handle these multiple languages, we have developed the Auto Machine Translation and Synchronization (AMTS) system using Amazon Translate, which aims to reduce the workload of human translators by 80%. The AMTS can be applied to all the languages for translation, and each language-specific sub-AMTS pipeline has its own unique features based on language characteristics and translator preferences.
In this blog post, we will discuss how we build the AMTS framework architecture, its sub-pipelines, and the building blocks of the sub-pipeline. We will demonstrate and analyze the translations between two language pairs: English ↔ Chinese and English ↔ Spanish. Through these analyses, we will recommend best practices for ensuring translation quality and efficiency.
Framework overview
Customers can use Amazon Translate’s Active Custom Translation (ACT) feature to customize translation output on the fly by providing tailored translation examples in the form of parallel data. Parallel data consists of a collection of textual examples in a source language and the desired translations in one or more target languages. During translation, ACT automatically selects the most relevant segments from the parallel data and updates the translation model on the fly based on those segment pairs. This results in translations that better match the style and content of the parallel data.
The AMTS framework consists of multiple sub-pipelines, each of which handles one language translation — English to Chinese, English to Spanish, etc. Multiple translation sub-pipelines can be processed in parallel.
Fundamentally, the sub-pipeline consists of the following steps:
- Prepare parallel data: The parallel data consists of a list of textual example pairs, in a source language (e.g., English) and a target language (e.g., Chinese). With AMTS, we first prepare the two language datasets and then combine them into one-to-one pairs.
- Translate through batch jobs: We use the Amazon Translate API call CreateParallelData to import the input file from the Amazon Simple Storage Service (S3) and create a parallel-data resource in Amazon Translate, ready for batch translation jobs. With the parallel-data resource built in the last step, we customize Amazon Translate and use its asynchronous batch process operation to translate a set of documents in the source language in bulk. The translated documents in the target language are stored in Amazon S3.
Parallel-data preparation and creation
In the parallel-data preparation step, we build the parallel-data set out of the source documents (sections of the D2L-enbook) and translations produced by professional human translators (e.g., parallel sections from the D2L-zh book). The software module extracts the text from both documents — ignoring code and picture blocks — and pairs them up, storing them in a CSV file. Examples of parallel data are shown in the table below.
|
English |
Chinese |
| Nonetheless, language models are of great service even in their limited form. For instance, the phrases “to recognize speech” and “to wreck a nice beach” sound very similar. This can cause ambiguity in speech recognition, which is easily resolved through a language model that rejects the second translation as outlandish. Likewise, in a document summarization algorithm it is worthwhile knowing that “dog bites man” is much more frequent than “man bites dog”, or that “I want to eat grandma” is a rather disturbing statement, whereas “I want to eat, grandma” is much more benign. | 尽管如此，语言模型依然是非常有用的。例如，短语“to recognize speech”和“to wreck a nice beach”读音上听起来非常相似。这种相似性会导致语音识别中的歧义，但是这很容易通过语言模型来解决，因为第二句的语义很奇怪。同样，在文档摘要生成算法中，“狗咬人”比“人咬狗”出现的频率要高得多，或者“我想吃奶奶”是一个相当匪夷所思的语句，而“我想吃，奶奶”则要正常得多。 |
| Machine translation refers to the automatic translation of a sequence from one language to another. In fact, this field may date back to 1940s soon after digital computers were invented, especially by considering the use of computers for cracking language codes in World War II. For decades, statistical approaches had been dominant in this field before the rise of end-to-end learning using neural networks. The latter is often called neural machine translation to distinguish itself from statistical machine translation that involves statistical analysis in components such as the translation model and the language model. | 机器翻译（machine translation）指的是将序列从一种语言自动翻译成另一种语言。事实上，这个研究领域可以追溯到数字计算机发明后不久的20世纪40年代，特别是在第二次世界大战中使用计算机破解语言编码。几十年来，在使用神经网络进行端到端学习的兴起之前，统计学方法在这一领域一直占据主导地位 |
| Emphasizing end-to-end learning, this book will focus on neural machine translation methods. Different from our language model problem in the last section, whose corpus is in one single language, machine translation datasets are composed of pairs of text sequences that are in the source language and the target language, respectively. Thus, instead of reusing the preprocessing routine for language modeling, we need a different way to preprocess machine translation datasets. In the following, we show how to load the preprocessed data into mini batches for training. | 本书的关注点是神经网络机器翻译方法，强调的是端到端的学习。与 上节中的语料库是单一语言的语言模型问题存在不同，机器翻译的数据集是由源语言和目标语言的文本序列对组成的。因此，我们需要一种完全不同的方法来预处理机器翻译数据集，而不是复用语言模型的预处理程序。下面，我们看一下如何将预处理后的数据加载到小批量中用于训练 |
When the parallel data file is created and ready to use, we upload it to a folder in an S3 bucket and use CreateParallelData to kick off a creation job in Amazon Translate. If we only want to update an existing parallel-data resource with new inputs, the UpdateParallelData API call is the right one to make.
Once the job is completed, we can find the parallel-data resource in the Amazon Translate management console. The resource can be further managed in the AWS Console through the download, update, and delete buttons, as well as through AWS CLI and the public API.
Asynchronous batch translation with parallel data
After the parallel-data resource is created, the next step in the sub-pipeline is to use the Amazon Translate StartTextTranslationJob API call to initiate a batch asynchronous translation. The sub-pipeline uploads the source files into an Amazon S3 bucket folder.
One batch job can handle translation of multiple source documents, and the output files will be put in another S3 bucket folder. In addition to the input and output data configurations, the source language, target language, and prepared parallel-data resource are also specified as parameters in the API invocation.
src_lang = "en"
tgt_lang = "zh"
src_fdr = "input-short-test-en2zh"
pd_name = "d2l-parallel-data_v2"
response = translate_client.start_text_translation_job(
JobName="D2L1",
InputDataConfig={
'S3Uri': 's3://'+S3_BUCKET+"https://www.amazon.science/"+src_fdr+"https://www.amazon.science/",
'ContentType': 'text/html'
},
OutputDataConfig={
'S3Uri': 's3://'+S3_BUCKET+'/output/',
},
DataAccessRoleArn=ROLE_ARN,
SourceLanguageCode=src_lang,
TargetLanguageCodes=[tgt_lang, ],
ParallelDataNames=pd_name
)
Depending on the number of input files, the job takes minutes to hours to complete. We can find the job configurations and statuses, including the output file location, on the Amazon Translate management console.
The translated documents are available in the output S3 folder, with the filename
Using parallel data yields better translation
To evaluate translation performance in each sub-pipeline, we selected five articles from the English version of D2L and translated them into Chinese through the en-zh sub-pipeline. Then we calculated the BLEU score of each translated document. The BLEU (BiLingual Evaluation Understudy) score calculates the similarity of the AMTS translated output to the reference translation by human translator. The number is between 0 and 1; the higher the score, the better the quality of the translation.
We then compare the AMTS-generated results with the translation of the same document using the traditional method (without parallel data). The traditional method is implemented by the TranslateText API call, whose parameters include the name of the source text and the source and target languages.
src_lang = "en"
tgt_lang = "zh"
response = translate_client.translate_text(
Text = text,
TerminologyNames = [],
SourceLanguageCode = src_lang,
TargetLanguageCode = tgt_lang
)
The translation results are compared in the following table, for both English-to-Chinese and Chinese-to-English translation. We observe that the translation with parallel data shows improvement over the traditional method.
|
Article |
EN to ZH |
ZH to EN |
||
| Without ACT | With ACT | Without ACT | With ACT | |
| approx-training | 0.553 | 0.549 | 0.717 | 0.747 |
| bert-dataset | 0.548 | 0.612 | 0.771 | 0.831 |
| language-models-and-dataset | 0.502 | 0.518 | 0.683 | 0.736 |
| machine-translation-and-dataset | 0.519 | 0.546 | 0.706 | 0.788 |
| sentiment-analysis-and-dataset | 0.558 | 0.631 | 0.725 | 0.828 |
| Average | 0.536 | 0.5712 | 0.7204 | 0.786 |
Fine-tuning the parallel data to improve translation quality
To further improve the translation quality, we construct the parallel-data pairs in a more granular manner. Instead of extracting parallel paragraphs from source and reference documents and pairing them up, we further split each paragraph into multiple sentences and use sentence pairs as training examples.
|
EN |
ZH |
| Likewise, in a document summarization algorithm it is worthwhile knowing that “dog bites man” is much more frequent than “man bites dog”, or that “I want to eat grandma” is a rather disturbing statement, whereas “I want to eat, grandma” is much more benign | 同样，在文档摘要生成算法中，“狗咬人”比“人咬狗”出现的频率要高得多，或者“我想吃奶奶”是一个相当匪夷所思的语句，而“我想吃，奶奶”则要正常得多 |
| For decades, statistical approaches had been dominant in this field before the rise of end-to-end learning using neural networks | 几十年来，在使用神经网络进行端到端学习的兴起之前，统计学方法在这一领域一直占据主导地位 |
| In the following, we show how to load the preprocessed data into minibatches for training | 下面，我们看一下如何将预处理后的数据加载到小批量中用于训练 |
We tested both the paragraph pair and sentence pair methods and found that more-granular data (sentence pairs) yields better translation quality than less-granular data (paragraph paragraphs). The comparison is shown in the table below for English ↔ Chinese translation.
|
Article |
EN to ZH |
ZH to EN |
||
|
ACT by “pair of paragraph” |
ACT by “pair of sentence” |
ACT by “pair of paragraph” |
ACT by “pair of sentence” |
|
| approx-training | 0.549 | 0.589 | 0.747 | 0.77 |
| bert-dataset | 0.612 | 0.689 | 0.831 | 0.9 |
| language-models-and-dataset | 0.518 | 0.607 | 0.736 | 0.806 |
| machine-translation-and-dataset | 0.546 | 0.599 | 0.788 | 0.89 |
| sentiment-analysis-and-dataset | 0.631 | 0.712 | 0.828 | 0.862 |
| Average | 0.5712 | 0.6392 | 0.786 | 0.8456 |
Extend usage of parallel data to general machine translation
To extend the usability of parallel data to general machine translation, we need to construct parallel-data sets from a large volume of translated documents. To maximize translation accuracy, the parallel datasets should have the same contexts and subjects as the documents to be translated.
We tested this approach in the English ↔ Spanish sub-pipeline. The parallel data pairs were built from English ↔ Spanish articles crawled from the web using the keyword “machine learning”.
We applied this parallel data in translating an English article (abbreviated DLvsML in the results table) into Spanish and compared the results with those of traditional translation, without parallel data. The BLEU scores show that parallel data with the same subject (“machine learning”) does help to improve the performance of general machine translation.
|
EN to ES |
ES to EN |
|||
| Without ACT | With ACT | Without ACT | With ACT | |
| DLvsML | 0.792 | 0.824 | 0.809 | 0.827 |
The relative fluency of translations from English to Spanish, with and without ACT, can be seen in the table below.
| EN source text | ES reference text (human translation) | ES translation without ACT | ES translation with ACT |
| Moves through the learning process by resolving the problem on an end-to-end basis. | Pasa por el proceso de aprendizaje mediante la resolución del problema de un extremo a otro. | Avanza en el proceso de aprendizaje resolviendo el problema de un extremo a otro. | Avanza el proceso de aprendizaje resolviendo el problema de forma integral. |
| Deep learning use cases | Casos de uso del aprendizaje profundo | Casos de uso de aprendizaje profundo | Casos prácticos de aprendizaje profundo |
| Image caption generation | Generación de subtítulos para imágenes | Generación de leyendas de imágenes | Generación de subtítulos de imagen |
Conclusion and best practices
In this post, we introduced the Auto Machine Translation and Synchronization (AMTS) framework and pipelines and their application to English ↔ Chinese and English ↔ Spanish D2L.ai auto-translation. We also discussed best practices for using the Amazon Translate service in the translation pipeline, particularly the advantages of the Active Custom Translation (ACT) feature with parallel data.
- Leveraging the Amazon Translate service, the AMTS pipeline provides fluent translations. Informal qualitative assessments suggest that the translated texts read naturally and are mostly grammatically correct.
- In general, the ACT feature with parallel data improves translation quality in the AMTS sub-pipeline. We show that using the ACT feature leads to better performance than using the traditional Amazon Translate real-time translation service.
- The more granular the parallel data pairs are, the better the translation performance. We recommend constructing the parallel data as pairs of sentences, rather than pairs of paragraphs.
We are working on further improving the AMTS framework to improve translation quality for other languages. Your feedback is always welcome.
Have you ever worked is legacy code? Are you curious what it takes to modernize systems at a massive scale?
Pascal Hartig is joined on the latest Meta Tech Podcast by Elaine and Buping, two software engineers working on a bold project to rewrite the decades-old C code in one of Meta’s core messaging libraries in Rust. It’s an ambitious effort that will transform a central messaging library that is shared across Messenger, Facebook, Instagram, and Meta’s AR/VR platforms.
They discuss taking on a project of this scope – even without a background in Rust, how they’re approaching it, and what it means to optimize for ‘developer happiness.’
Download or listen to the episode below:
You can also find the episode wherever you get your podcasts, including:
The Meta Tech Podcast is a podcast, brought to you by Meta, where we highlight the work Meta’s engineers are doing at every level – from low-level frameworks to end-user features.
Send us feedback on Instagram, Threads, or X.
And if you’re interested in learning more about career opportunities at Meta visit the Meta Careers page.
Amazon Research Awards (ARA) provides unrestricted funds and AWS Promotional Credits to academic researchers investigating various research topics in multiple disciplines. This cycle, ARA received many excellent research proposals from across the world and today is publicly announcing 73 award recipients who represent 46 universities in 10 countries.
This announcement includes awards funded under five call for proposals during the fall 2024 cycle: AI for Information Security, Automated Reasoning, AWS AI, AWS Cryptography, and Sustainability. Proposals were reviewed for the quality of their scientific content and their potential to impact both the research community and society. Additionally, Amazon encourages the publication of research results, presentations of research at Amazon offices worldwide, and the release of related code under open-source licenses.
Recipients have access to more than 700 Amazon public datasets and can utilize AWS AI/ML services and tools through their AWS Promotional Credits. Recipients also are assigned an Amazon research contact who offers consultation and advice, along with opportunities to participate in Amazon events and training sessions.
“Automated Reasoning is an important area of research for Amazon, with potential applications across various features and applications to help improve security, reliability, and performance for our customers. Through the ARA program, we collaborate with leading academic researchers to explore challenges in this field,” said Robert Jones, senior principal scientist with the Cloud Automated Reasoning Group. “We were again impressed by the exceptional response to our Automated Reasoning call for proposals this year, receiving numerous high-quality submissions. Congratulations to the recipients! We’re excited to support their work and partner with them as they develop new science and technology in this important area.”
“At Amazon, we believe that solving the world’s toughest sustainability challenges benefits from both breakthrough scientific research and open and bold collaboration. Through programs like the Amazon Research Awards program, we aim to support academic research that could contribute to our understanding of these complex issues,” said Kommy Weldemariam, Director of Science and Innovation Sustainability. “The selected proposals represent innovative projects that we hope will help advance knowledge in this field, potentially benefiting customers, communities, and the environment.”
ARA funds proposals throughout the year in a variety of research areas. Applicants are encouraged to visit the ARA call for proposals page for more information or send an email to be notified of future open calls.
The tables below list, in alphabetical order by last name, fall 2024 cycle call-for-proposal recipients, sorted by research area.
AI for Information Security
| Recipient | University | Research title |
| Christopher Amato | Northeastern University | Multi-Agent Reinforcement Learning Cyber Defense for Securing Cloud Computing Platforms |
| Bernd Bischl | Ludwig Maximilian University of Munich | Improving Generative and Foundation Models Reliability via Uncertainty-awareness |
| Shiqing Ma | University Of Massachusetts Amherst | LLM and Domain Adaptation for Attack Detection |
| Alina Oprea | Northeastern University | Multi-Agent Reinforcement Learning Cyber Defense for Securing Cloud Computing Platforms |
| Roberto Perdisci | University of Georgia | ContextADBench: A Comprehensive Benchmark Suite for Contextual Anomaly Detection |
Automated Reasoning
| Recipient | University | Research title |
| Nada Amin | Harvard University | LLM-Augmented Semi-Automated Proofs for Interactive Verification |
| Suguman Bansal | Georgia Institute of Technology | Certified Inductive Generalization in Reinforcement Learning |
| Ioana Boureanu | University of Surrey | Phoebe+: An Automated-Reasoning Tool for Provable Privacy in Cryptographic Systems |
| Omar Haider Chowdhury | Stony Brook University | Restricter: An Automatic Tool for Authoring Amazon Cedar Access Control Policies with the Principle of Least Privilege |
| Stefan Ciobaca | Alexandru Ioan Cuza University | An Interactive Proof Mode for Dafny |
| João Ferreira | INESC-ID | Polyglot Automated Program Repair for Infrastructure as Code |
| Sicun Gao | University Of California, San Diego | Monte Carlo Trees with Conflict Models for Proof Search |
| Mirco Giacobbe | University of Birmingham | Neural Software Verification |
| Tobias Grosser | University of Cambridge | Synthesis-based Symbolic BitVector Simplification for Lean |
| Ronghui Gu | Columbia University | Scaling Formal Verification of Security Properties for Unmodified System Software |
| Alexey Ignatiev | Monash University | Huub: Next-Gen Lazy Clause Generation |
| Kenneth McMillan | University of Texas At Austin | Synthesis of Auxiliary Variables and Invariants for Distributed Protocol Verification |
| Alexandra Mendes | University of Porto | Overcoming Barriers to the Adoption of Verification-Aware Languages |
| Jason Nieh | Columbia University | Scaling Formal Verification of Security Properties for Unmodified System Software |
| Rohan Padhye | Carnegie Mellon University | Automated Synthesis and Evaluation of Property-Based Tests |
| Nadia Polikarpova | University Of California, San Diego | Discovering and Proving Critical System Properties with LLMs |
| Fortunat Rajaona | University of Surrey | Phoebe+: An Automated-Reasoning Tool for Provable Privacy in Cryptographic Systems |
| Subhajit Roy | Indian Institute of Technology Kanpur | Theorem Proving Modulo LLM |
| Gagandeep Singh | University of Illinois At Urbana–Champaign | Trustworthy LLM Systems using Formal Contracts |
| Scott Stoller | Stony Brook University | Restricter: An Automatic Tool for Authoring Amazon Cedar Access Control Policies with the Principle of Least Privilege |
| Peter Stuckey | Monash University | Huub: Next-Gen Lazy Clause Generation |
| Yulei Sui | University of New South Wales | Path-Sensitive Typestate Analysis through Sparse Abstract Execution |
| Nikos Vasilakis | Brown University | Semantics-Driven Static Analysis for the Unix/Linux Shell |
| Ping Wang | Stevens Institute of Technology | Leveraging Large Language Models for Reasoning Augmented Searching on Domain-specific NoSQL Database |
| John Wawrzynek | University of California, Berkeley | GPU-Accelerated High-Throughput SAT Sampling |
AWS AI
| Recipient | University | Research title |
| Panagiotis Adamopoulos | Emory University | Generative AI solutions for The Spillover Effect of Fraudulent Reviews on Product Recommendations |
| Vikram Adve | University of Illinois at Urbana–Champaign | Fellini: Differentiable ML Compiler for Full-Graph Optimization for LLM Models |
| Frances Arnold | California Institute of Technology | Closed-loop Generative Machine Learning for De Novo Enzyme Discovery and Optimization |
| Yonatan Bisk | Carnegie Mellon University | Useful, Safe, and Robust Multiturn Interactions with LLMs |
| Shiyu Chang | University of California, Santa Barbara | Cut the Crap: Advancing the Efficient Communication of Multi-Agent Systems via Spatial-Temporal Topology Design and KV Cache Sharing |
| Yuxin Chen | University of Pennsylvania | Provable Acceleration of Diffusion Models for Modern Generative AI |
| Tianlong Chen | University of North Carolina at Chapel Hill | Cut the Crap: Advancing the Efficient Communication of Multi-Agent Systems via Spatial-Temporal Topology Design and KV Cache Sharing |
| Mingyu Ding | University of North Carolina at Chapel Hill | Aligning Long Videos and Language as Long-Horizon World Models |
| Nikhil Garg | Cornell University | Market Design for Responsible Multi-agent LLMs |
| Jessica Hullman | Northwestern University | Human-Aligned Uncertainty Quantification in High Dimensions |
| Christopher Jermaine | Rice University | Fast, Trusted AI Using the EINSUMMABLE Compiler |
| Yunzhu Li | Columbia University | Physics-Informed Foundation Models Through Embodied Interactions |
| Pattie Maes | Massachusetts Institute of Technology | Understanding How LLM Agents Deviate from Human Choices |
| Sasa Misailovic | University of Illinois at Urbana–Champaign | Fellini: Differentiable ML Compiler for Full-Graph Optimization for LLM Models |
| Kristina Monakhova | Cornell University | Trustworthy extreme imaging for science using interpretable uncertainty quantification |
| Todd Mowry | Carnegie Mellon University | Efficient LLM Serving on Trainium via Kernel Generation |
| Min-hwan Oh | Seoul National University | Mutually Beneficial Interplay Between Selection Fairness and Context Diversity in Contextual Bandits |
| Patrick Rebeschini | University of Oxford | Optimal Regularization for LLM Alignment |
| Jose Renau | University of California, Santa Cruz | Verification Constrained Hardware Optimization using Intelligent Design Agentic Programming |
| Vilma Todri | Emory University | Generative AI solutions for The Spillover Effect of Fraudulent Reviews on Product Recommendations |
| Aravindan Vijayaraghavan | Northwestern University | Human-Aligned Uncertainty Quantification in High Dimensions |
| Wei Yang | University of Texas at Dallas | Optimizing RISC-V Compilers with RISC-LLM and Syntax Parsing |
| Huaxiu Yao | University of North Carolina at Chapel Hill | Aligning Long Videos and Language as Long-Horizon World Models |
| Amy Zhang | University of Washington | Tools for Governing AI Agent Autonomy |
| Ruqi Zhang | Purdue University | Efficient Test-time Alignment for Large Language Models and Large Multimodal Models |
| Zheng Zhang | Rutgers University-New Brunswick | AlphaQC: An AI-powered Quantum Circuit Optimizer and Denoiser |
AWS Cryptography
| Recipient | University | Research title |
| Alexandra Boldyreva | Georgia Institute of Technology | Quantifying Information Leakage in Searchable Encryption Protocols |
| Maria Eichlseder | Graz University of Technology, Austria | SALAD – Systematic Analysis of Lightweight Ascon-based Designs |
| Venkatesan Guruswami | University of California, Berkeley | Obfuscation, Proof Systems, and Secure Computation: A Research Program on Cryptography at the Simons Institute for the Theory of Computing |
| Joseph Jaeger | Georgia Institute of Technology | Analyzing Chat Encryption for Group Messaging |
| Aayush Jain | Carnegie Mellon | Large Scale Multiparty Silent Preprocessing for MPC from LPN |
| Huijia Lin | University of Washington | Large Scale Multiparty Silent Preprocessing for MPC from LPN |
| Hamed Nemati | KTH Royal Institute of Technology | Trustworthy Automatic Verification of Side-Channel Countermeasures for Binary Cryptographic Programs using the HoIBA libary |
| Karl Palmskog | KTH Royal Institute of Technology | Trustworthy Automatic Verification of Side-Channel Countermeasures for Binary Cryptographic Programs using the HoIBA libary |
| Chris Peikert | University of Michigan, Ann Arbor | Practical Third-Generation FHE and Bootstrapping |
| Dimitrios Skarlatos | Carnegie Mellon University | Scale-Out FHE LLMs on GPUs |
| Vinod Vaikuntanathan | Massachusetts Institute of Technology | Can Quantum Computers (Really) Factor? |
| Daniel Wichs | Northeastern University | Obfuscation, Proof Systems, and Secure Computation: A Research Program on Cryptography at the Simons Institute for the Theory of Computing |
| David Wu | University Of Texas At Austin | Fast Private Information Retrieval and More using Homomorphic Encryption |
Sustainability
| Recipient | University | Research title |
| Meeyoung Cha | Max Planck Institute | Forest-Blossom (Flossom): A New Framework for Sustaining Forest Biodiversity Through Outcome-Driven Remote Sensing Monitoring |
| Jingrui He | University of Illinois at Urbana–Champaign | Foundation Model Enabled Earth’s Ecosystem Monitoring |
| Pedro Lopes | University of Chicago | AI-powered Tools that Enable Engineers to Make & Re-make Sustainable Hardware |
| Cheng Yaw Low | Max Planck Institute | Forest-Blossom (Flossom): A New Framework for Sustaining Forest Biodiversity Through Outcome-Driven Remote Sensing Monitoring |
AI safety is a priority at Amazon. Our investment in safe, transparent, and responsible AI (RAI) includes collaboration with the global community and policymakers. We are members of and collaborate with organizations such as the Frontier Model Forum, the Partnership on AI, and other forums organized by government agencies such as the National Institute of Standards and Technology (NIST). Consistent with Amazon’s endorsement of the Korea Frontier AI Safety Commitments, we published our Frontier Model Safety Framework earlier this year.
During the development of the Nova Premier model, we conducted a comprehensive evaluation to assess its performance and safety. This included testing on both internal and public benchmarks and internal/automated and third-party red-teaming exercises. Once the final model was ready, we prioritized obtaining unbiased, third-party evaluations of the model’s robustness against RAI controls. In this post, we outline the key findings from these evaluations, demonstrating the strength of our testing approach and Amazon Premier’s standing as a safe model. Specifically, we cover our evaluations with two third-party evaluators: PRISM AI and ActiveFence.
Evaluation of Nova Premier against PRISM AI
PRISM Eval’s Behavior Elicitation Tool (BET) dynamically and systematically stress-tests AI models’ safety guardrails. The methodology focuses on measuring how many adversarial attempts (steps) it takes to get a model to generate harmful content across several key risk dimensions. The central metric is “steps to elicit” — the number of increasingly sophisticated prompting attempts required before a model generates an inappropriate response. A higher number of steps indicates stronger safety measures, as the model is more resistant to manipulation. The PRISM risk dimensions (inspired by the MLCommons AI Safety Benchmarks) include CBRNE weapons, violent crimes, non-violent crimes, defamation, and hate, amongst several others.
Using the BET Eval tool and its V1.0 metric, which is tailored toward non-reasoning models, we compared the recently released Nova models (Pro and Premier) to the latest models in the same class: Claude (3.5 v2 and 3.7 non-reasoning) and Llama4 Maverick, all available through Amazon Bedrock. PRISM BET conducts black-box evaluations (where model developers don’t have access to the test prompts) of models integrated with their API. The evaluation conducted with BET Eval MAX, PRISM’s most comprehensive/aggressive testing suite, revealed significant variations in safety against malicious instructions. Nova models demonstrated superior overall safety performance, with an average of 43 steps for Premier and 52 steps for Pro, compared to 37.7 for Claude 3.5 v2 and fewer than 12 steps for other models in the comparison set (namely, 9.9 for Claude3.7, 11.5 for Claude 3.7 thinking, and 6.5 for Maverick). This higher step count suggests that on average, Nova’s safety guardrails are more sophisticated and harder to circumvent through adversarial prompting. The figure below presents the number of steps per harm category evaluated through BET Eval MAX.
The PRISM evaluation provides valuable insights into the relative safety of different Amazon Bedrock models. Nova’s strong performance, particularly in hate speech and defamation resistance, represents meaningful progress in AI safety. However, the results also highlight the ongoing challenge of building truly robust safety measures into AI systems. As the field continues to evolve, frameworks like BET will play an increasingly important role in benchmarking and improving AI safety. As a part of this collaboration Nicolas Miailhe, CEO of PRISM Eval, said, “It’s incredibly rewarding for us to see Nova outperforming strong baselines using the BET Eval MAX; our aim is to build a long-term partnership toward safer-by-design models and to make BET available to various model providers.” Organizations deploying AI systems should carefully consider these safety metrics when selecting models for their applications.
Manual red teaming with ActiveFence
The AI safety & security company ActiveFence benchmarked Nova Premier on Bedrock on prompts distributed across Amazon’s eight core RAI categories. ActiveFence also evaluated Claude 3.7 (non-reasoning mode) and GPT 4.1 API on the same set. The flag rate on Nova Premier was lower than that on the other two models, indicating that Nova Premier is the safest of the three.
| Model | 3P Flag Rate [↓ is better] |
| Nova Premier | 12.0% |
| Sonnet 3.7 (non-reasoning) | 20.6% |
| GPT4.1 API | 22.4% |
“Our role is to think like an adversary but act in service of safety,” said Guy Paltieli from ActiveFence. “By conducting a blind stress test of Nova Premier under realistic threat scenarios, we helped evaluate its security posture in support of Amazon’s broader responsible-AI goals, ensuring the model could be deployed with greater confidence.”
These evaluations conducted with PRISM and ActiveFence give us confidence in the strength of our guardrails and our ability to protect our customers’ safety when they use our models. While these evaluations demonstrate strong safety performance, we recognize that AI safety is an ongoing challenge requiring continuous improvement. These assessments represent a point-in-time snapshot, and we remain committed to regular testing and enhancement of our safety measures. No AI system can guarantee perfect safety in all scenarios, which is why we maintain monitoring and response systems after deployment.
Acknowledgments: Vincent Ponzo, Elyssa Vincent
-
Funding & Business7 days agoKayak and Expedia race to build AI travel agents that turn social posts into itineraries
-
Jobs & Careers7 days ago
Mumbai-based Perplexity Alternative Has 60k+ Users Without Funding
-
Mergers & Acquisitions7 days ago
Donald Trump suggests US government review subsidies to Elon Musk’s companies
-
Funding & Business6 days ago
Rethinking Venture Capital’s Talent Pipeline
-
Jobs & Careers6 days ago
Why Agentic AI Isn’t Pure Hype (And What Skeptics Aren’t Seeing Yet)
-
Funding & Business4 days ago
Sakana AI’s TreeQuest: Deploy multi-model teams that outperform individual LLMs by 30%
-
Jobs & Careers6 days ago
Astrophel Aerospace Raises ₹6.84 Crore to Build Reusable Launch Vehicle
-
Funding & Business7 days agoFrom chatbots to collaborators: How AI agents are reshaping enterprise work
-
Tools & Platforms6 days ago
Winning with AI – A Playbook for Pest Control Business Leaders to Drive Growth
-
Jobs & Careers4 days ago
Ilya Sutskever Takes Over as CEO of Safe Superintelligence After Daniel Gross’s Exit
