December 2024 Summaries

The Gretel's Synthetic Safety Dataset is a resource designed to align large language models (LLMs) with safe and ethical responses. The dataset features 8,361 triplets of "prompt", "response" and "safe response" spanning significant risk categories, including discrimination, harassment, propaganda, religious intolerance, gender bias, and more. It was created using Gretel Navigator's Data Designer toolkit and is available on HuggingFace. The dataset aims to provide a transparent and modular resource for the AI community to utilize in aligning models for secure and public-interest-focused interactions. It also highlights the importance of prompt generation benefits from human expertise in jailbreaking (attempts to bypass model restrictions) and red teaming (simulated attacks to test system security). The dataset can be used for pre-training and fine-tuning guardrails, stress-testing model robustness, facilitating rapid iteration and refinement, and benchmarking ethical and safety maturity.

Gretel's latest innovation is Model Suites for synthetic data generation, which simplifies the process of creating high-quality synthetic data while addressing critical challenges like licensing and compliance. The future of generative AI applications and services will be compound, integrating multiple tools and models working together in a coordinated way, rather than relying on single large language models. Compound AI systems are designed to tackle complex tasks using multiple interacting components, including multiple calls to models, retrievers, or external tools. Gretel's Navigator is a first compound AI system purpose-built for generating and iterating on synthetic data, providing state-of-the-art (SOTA) tools to help teams eliminate bottlenecks. Model Suites offer curated collections of tools and models designed to simplify model selection, ensure license compliance, guarantee quality, offer flexibility, provide transparency, and make it easy for users to navigate complex licensing complexities while maintaining flexibility, compliance, and high-quality synthetic data generation.

Gretel Navigator is now available on Amazon Bedrock, offering a solution for generating synthetic training data needed to connect language models with real-world tools and APIs. The biggest challenge in building AI systems for computer control is the training data, which often lacks important edge cases and error scenarios. Gretel Navigator on Amazon Bedrock provides enterprise-grade infrastructure, streamlined workflow, cost-effective synthetic data generation, and enhanced data quality to address these challenges. This integration offers a scalable, efficient path to better training data for AI applications such as function calling capabilities, tool use scenarios, and more.