@repligate This paper shows models can verbatim memorize data from RL, especially from DPO/IPO (~similar memorization to SFT at ~18%) but also specifically prompts from PPO (at ~0.4%, which is notably not 0%) https://t.co/HeIaW9MxWF
in reply to: 1994973338448662858
same thread: 1994973338448662858 1994984081906700757 1994997440525865353 1994997903824437598 1995006276825461244 1995008030317170819 1995008116635931109 1995016497375449289 1995016961617707058 1995017859458896103 1995021640560959536 1995028164079489295 1995053425030222125 1995053962396152289 1995054216738705645 1995074341739098246 1995323510449922268 1995325660278247816 1995664107857739830 1995664966641234390
Reproduced against link rot, credited and linked to its original. Yours and you’d rather it weren’t here? Open an issue.