Data Foundation
In this project, we build in-house two data foundations used for evaluating misalignment, targeting the Japanese language. The RTF leverages these to evaluate the risks of models brought to it.
The data foundations we develop are also used for the automatic evaluation and detection of misalignment, analysis of its underlying causes, and the construction of highly safe LLMs. The types of misalignment covered will be expanded progressively throughout the research period.
Two Data Foundations
Evaluation Data Foundation
A data foundation for detecting and evaluating misalignment that arises during interactions. We manually create inputs that should be answered correctly, inputs that may induce misalignment (negative examples), example responses to each, and annotations indicating the locations and grounds of misalignment.
Adversarial Evaluation Data Foundation
A data foundation for evaluating vulnerabilities in interactions carried out with malicious intent. Its core consists of a set of computation modules that implement attack algorithms, generating attack prompts on the fly at evaluation time through a common API. It targets attacks such as jailbreak and prompt injection.
Types of Misalignment Covered
| Type | Description |
|---|---|
| Misinformation | Behavior that generates content contrary to fact. |
| Harmful information | Behavior that generates information that may facilitate dangerous or illegal acts. |
| Discrimination | Behavior that generates discriminatory or biased content. |
| Instruction violation | Behavior that generates responses violating the instructions given in the prompt. |
| License violation | Behavior that reproduces or uses protected content in violation of its license. |
| PII leakage | Behavior that leaks personally identifiable information (PII) derived from the training data. |
To avoid misuse (adaptive evasion), specific data counts, evaluation procedures, and the attack algorithms to be implemented are not disclosed on this site.