I think steering is a good idea for getting diversity of responses in post-training, where the diverse responses are then ranked for DPO/GRPO by a judge model that is not steered.
in reply to: 2051647981485470197
same thread: 2051647981485470197 2051647986120180225 2051647989001707632 2051647997990117596 2051698011240058939 2051699080644935778 2051699730493657369 2051790135918768437
Reproduced against link rot, credited and linked to its original. Yours and you’d rather it weren’t here? Open an issue.