Certainly! Here’s the translation: “Discover Custom Models on Amazon Bedrock: Unlock New Perspectives with Logarithmic Probability Support.”

Amazon Bedrock, the company’s cloud artificial intelligence platform, has implemented a new feature that provides support for log probabilities in its Custom Model Import function. This addition allows developers to integrate models such as Llama, Mistral, and Qwen, which have been fine-tuned on other platforms, thus offering a serverless experience and a unified API.

The ability to understand a model’s confidence in its predictions is crucial for creating reliable artificial intelligence applications. This is particularly relevant when working with custom models that respond to specific queries across various domains. With the new support for log probabilities, users can now access detailed information about the model’s certainty for each prediction at the token level, enhancing visibility into the model’s behavior and enabling more effective evaluation and filtering of results.

Log probabilities make it easier for developers to measure the model’s confidence regarding each generated token. Values closer to zero indicate higher confidence; for example, a value of -0.1 corresponds to approximately 90% certainty, in contrast to -3.0, which translates to around 5% confidence. This new capability allows users to examine certainty in different parts of a response, identify potential errors, and optimize costs in retrieval-augmented generation (RAG) systems.

To take advantage of these log probabilities in Custom Model Import, users must have an active AWS account that includes access to Amazon Bedrock, along with the necessary permissions to invoke models through the corresponding API. This advancement represents a step toward greater transparency in the decision-making process within models and empowers developers to create applications that adapt to varying levels of certainty, which is vital in critical fields such as healthcare and finance.

The applications of this functionality are numerous: from detecting hallucinations and low-confidence responses to optimizing RAG systems and evaluating model adjustments. By introducing quantifiable metrics on model confidence, Amazon Bedrock reinforces its commitment to providing tools that foster safe and scalable innovation in the field of artificial intelligence.