The NEREL dataset contains approximately 56,000 entities and 39,000 relations across 29 labeling types, derived from news articles. A local copy of NEREL-v1.0 includes 933 files with 55,950 entities and 38,673 relations.
The Russian portion of the MASSIVE dataset consists of 16,521 utterances. It is designed for intent recognition and slot filling rather than personal data identification, though it helps test short, informal queries.
The hivetrace/pii-bench dataset includes 1,810 examples covering 13 data types and 9 domains. It is specifically tailored for evaluating data masking gateways, though results depend on the specific recognition rules applied.
The AgentMask-RU project was created to link data masking with the restoration of values in tool calls. It tests both the protection of data and the agent's ability to execute tasks correctly after masking.
External datasets like ai4privacy, BFCL, τ-bench, and MERA are used to complement testing. However, they lack specific checks for the full chain of data masking and subsequent tool execution, which is the primary focus of AgentMask-RU.
