3 ms·
When I tried this, it made up random, non-repeatable confidence values that had little bearing to the veracity of its statements.
by ukuina 3y ago
When I tried this, it made up random, non-repeatable confidence values that had little bearing to the veracity of its statements.
- sebzim4500 3y agoIf you read the GPT-4 paper, they claim that the base model was well calibrated but the calibration was completely broken by RLHF