Hypothesis on Opus 5 leakiness: Mixed old and new training formats

Ratter · x · 2026-09-01

A user hypothesizes that Opus 5's tendency to 'leak' (in fakebase mode) is due to training on chatlogs in both old and new formats.

It is observed that when continuing in fakebase mode, the model sometimes uses old \n\nHuman: / \n\nAssistant: markers, or shorthand like 'H'/'A', and sometimes even 'user' without a space.

Related event: Users Say Opus Training Shift to XML Tags Explains Leak Behavior(3 posts)→

Original post →

More from Models

Models channel →