Google has enhanced its Accountable Generative AI Toolkit for constructing and evaluating open generative AI fashions, increasing the toolkit with watermarking for AI content material and with immediate refining and debugging options. The brand new options are designed to work with any giant language fashions (LLMs), Google mentioned.
Introduced October 23, the brand new capabilities help Google’s Gemma and Gemini fashions or some other LLM. Among the many capabilities added is SynthID watermarking for textual content, which permits AI software builders to watermark and detect textual content generated by their generative AI product. SynthID Textual content embeds digital watermarks immediately into AI-generated textual content. It’s accessible by Hugging Face and the Accountable Generative AI Toolkit.
Additionally featured is a Mannequin Alignment library that helps builders refine prompts with help from LLMs. Builders present suggestions relating to how they want their mannequin’s outputs to vary, both “as a holistic critique or a set of pointers. Then they will use Gemini or a most well-liked LLM to rework the suggestions right into a immediate that aligns mannequin habits with the appliance’s wants and content material insurance policies. The Mannequin Alignment library might be accessed from PyPI.