@harshad1313 i disagree, i suspect it balances the constraints 4.8 is under in rl / evals. i don't think it's straightforwardly 'hacking the grader,' and it's relatively shallow, like 4.8 learned the minimum amount of it necessary
same thread: 2061677549097484338 2061680656439218244 2061684889112445230 2061685146437509242 2061685460091736311 2061720486724444236 2061858368936165422 2061860996789293391 2061896547055640644 2061911848820322447 2062280631992930728 2062663685840224291 2062669401435087358 2062669839265939469
Reproduced against link rot, credited and linked to its original. Yours and you’d rather it weren’t here? Open an issue.