AWS ai documentation change
Summary
Fixed formatting issues in safety evaluation section and removed redundant numbers in Frontier AI Safety Commitments paragraph
Security assessment
Changes are editorial improvements to existing security documentation about model safety evaluations. No new security information or vulnerability fixes were introduced.
Diff
diff --git a/ai/responsible-ai/nova-micro-lite-pro/overview.md b/ai/responsible-ai/nova-micro-lite-pro/overview.md index 2304c8e1e..549c58610 100644 --- a//ai/responsible-ai/nova-micro-lite-pro/overview.md +++ b//ai/responsible-ai/nova-micro-lite-pro/overview.md @@ -23 +23 @@ The "overall effectiveness" of any traditional or generative model for a specifi -Amazon Nova FMs generate completions by understanding inputs and generating relevant completions, and, like more traditional ML solutions, must overcome issues of intrinsic and confounding variation. Intrinsic variation refers to features of the input to which the model should attend, for example, knowing the difference between the prompts "Did the cat win the game?" and "Did the dog win the game?" Confounding variation refers to features of the input that the model should ignore, for example, understanding that the prompts "Did my graceful feline friend win the game?" and "Did my cat win the game?" are asking about a cat, and not a dog. The full set of variations encountered by an FM includes language (human and machine), slang, professional jargon, dialects, expressive non-standard spelling and punctuation and many kinds of errors in prompts, for example, with spelling, grammar, punctuation, logic, and semantics. +Amazon Nova FMs generate completions by understanding inputs and generating relevant completions, and, like more traditional ML solutions, must overcome issues of intrinsic and confounding variation. Intrinsic variation refers to features of the input to which the model should attend, for example, knowing the difference between the prompts "Did the cat win the game?" and "Did the dog win the game?" Confounding variation refers to features friend win the game?" and "Did my cat win the game?" are asking about a cat, and not a dog. The full set of variations encountered by an FM includes language (human and machine), slang, professional jargon, dialects, expressive non-standard spelling and punctuation and many kinds of errors in prompts, for example, with spelling, grammar, punctuation, logic, and semantics. @@ -153 +153 @@ Safety is a shared responsibility between AWS and our customers. Our goal for sa - * _Frontier AI Safety Commitments:_ Consistent with our [Frontier AI Safety Commitments](https://www.amazon.science/publications/amazons-frontier-model-safety-framework), we comprehensively 270 evaluated Amazon Nova Premier — both its current public version and earlier iterations — across three 271 high-risk domains: Chemical, Biological, Radiological, and Nuclear (CBRN), Offensive Cyber Operations, and 272 Automated AI R&D. Using benchmarks and expert assessments, we found that Amazon Nova Premier is 273 within the tolerance threshold for critical capabilities in each domain. + * _Frontier AI Safety Commitments:_ Consistent with our [Frontier AI Safety Commitments](https://www.amazon.science/publications/amazons-frontier-model-safety-framework), we comprehensively evaluated Amazon Nova Premier — both its current public version and earlier iterations — across three high-risk domains: Chemical, Biological, Radiological, and Nuclear (CBRN), Offensive Cyber Operations, and Automated AI R&D. Using benchmarks and expert assessments, we found that Amazon Nova Premier is within the tolerance threshold for critical capabilities in each domain.