原书出处
Build a Reasoning Model (From Scratch)
keep = cumprobas <= top_p #C keep[:, 0] = True #D kept_sorted = torch.where( #E keep, sorted_probas, #E torch.zeros_like(sorted_probas) #E ) filtered = torch.zeros_like(probas).scatter(1, sorted_idx, kept_sorted) #F denom = torch.sum(filtered, dim=1).clamp_min(1e-12) #G return filtered / denom #A Step 4.1: Sort by descending probability #B Step 4.2: Cumulative sum #C Step 4.3.1: Keep tokens where cumprob <= top_p #D For top_p <= 0, only the highest-probability
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-4.html
EPUB SHA-256:52d4d07986020be70f5aac21762dfcf60fcba08db0ba29d4f63957caf12aa377