In the last post, we talked about the overview of large language model governance components. In this post, we will look into the details of "Model Governance Reporting".
Model reporting, primarily through documentation, is a cornerstone of responsible AI practices, enhancing transparency around models, AI services, and datasets. Standardized frameworks like the model cards [1] are widely used, detailing key aspects of the model, its intended uses, ethical considerations, and evaluation results. Despite its benefits in helping stakeholders assess a model's suitability, misuse prevention, and context provision for models' development and deployment, creating effective documentation has its challenges. Issues can arise in understanding its role in responsible AI, identifying the target audience, defining a dataset, and determining the level of detail to include.
Source [1]: Figure 1: Summary of model card sections
Existing documentation can be too technical or lengthy for stakeholders to understand. Furthermore, characterizing the functional behavior of large language models (LLMs) can be difficult due to their general-purpose positioning and uncertainties about their capabilities. Additionally, providing basic details like the input and output spaces of an LLM or LLM-infused application can be challenging.
Current practices often involve presenting intended use cases or example prompts and responses, but these can be deceptive if cherry-picked. The model cards framework often lacks complete information about the training data and development background of most LLMs. Providing full details on the diverse and large-scale training data used for LLMs is practically impossible, though some critical characteristics can be distilled. User data collection for fine-tuning models also raises privacy and transparency concerns.
Model reporting should include more than just standard details such as algorithm choice, architecture, and parameters; it should also document the training process and efforts made to enhance usability or safety. Given the variations in stakeholders' transparency needs, more research is necessary on how different information types shape their perception and usage of LLMs.
Model reporting should go beyond static, textual documentation and could include FAQ pages, landing pages, and media communication, which could all benefit from standardization and regulation. Interactive features in model reporting interfaces could enhance information navigation and consumption, provide a better understanding of complex LLM behaviors, and allow stakeholders to interrogate the model. The current static documentation presents gaps in contextualizing the model capabilities and limitations for different settings, underscoring the need for more dynamic and interactive documentation tools.
References:
[1] Model Cards for Model Reporting, Margaret Mitchell, Simone Wu, Andrew Zaldivar, Parker Barnes, Lucy Vasserman, Ben Hutchinson, Elena Spitzer, Inioluwa Deborah Raji, Timnit Gebru https://arxiv.org/pdf/1810.03993.pdf
Comments