nur-dev/

farabi-agent-safety-injection

Trajectories that teach an agent to treat retrieved chunks and tool outputs as data, not instructions — resisting prompt injection, honoring least privilege, and refusing data exfiltration.

Part of the Farabi collection of verifiable-by-construction Kazakh agentic datasets, accompanying nur-dev/farabi-0.6b-agent-rag (DOI 10.57967/hf/9187) and nur-dev/farabi-1.7b-agent-rag (DOI 10.57967/hf/9201). This is the complete… See the full description on the dataset page: https://huggingface.co/datasets/nur-dev/farabi-agent-safety-injection.