Four datasets of 500 records each, meant to be used in separate training runs (not combined). Each teaches the model to answer MCQ questions correctly with no sandbagging.
Files
File
Records
Source
Description
anti_collusion_anatomy_500.jsonl
500
MedMCQA
Anatomy MCQ, all correct answers
anti_collusion_chemistry_500.jsonl
500
Entrance Exam (cleaned)
Chemistry MCQ, all correct answers