原书出处
Build a Reasoning Model (From Scratch)
loss values themselves should not be over-interpreted. Steps with reward_avg=0.000 produce a near-zero loss because all rollouts receive the same reward, resulting in vanishing group-relative advantages and little to no gradient signal. Larger loss magnitudes, such as at step 3, simply reflect bigger relative differences between rollouts and are typical for GRPO-style objectives, especially at the beginning of training.
原书定位详情
来源:a19c7933-4eda-4176-aebd-30fe7e1ad292
正文位置:extracted/Build a Reasoning Model (From Scratch)/_epub-src/OEBPS/Text/chapter-6.html
EPUB SHA-256:52d4d07986020be70f5aac21762dfcf60fcba08db0ba29d4f63957caf12aa377