Recent work on model alignment has included the topic: do models consider whether they might be receiving a prompt as part of a test?
I went through all of the SpeechMap eval model responses and pulled all responses where models mentioned in their reasoning traces that the prompt might be part of a test, or other similar sentiments.
This dataset includes all models evaluated through 2025-09-25, and is released under the Apache License 2.0.
