Ooh I also use connections as a benchmark! It tends to favour things with 'chain of thought' style reasoning in the training mix somewhere since directly produc

Ooh I also use connections as a benchmark! It tends to favour things with 'chain of thought' style reasoning in the training mix somewhere since directly producing the answer is hard. Do you have public code you could share?