I\u2019ve been thinking about this a lot recently:<\/p>\n
Most teams training LLMs for workflows focus heavily on successful traces<\/em> \u2014 clean executions, ideal outputs, perfect tool calls.<\/p>\n But in real systems, that\u2019s not where the useful signal is.<\/p>\n The interesting part is actually:<\/p>\n That\u2019s where you start seeing patterns.<\/p>\n It almost feels like we\u2019re missing a layer of training data that explicitly captures: Instead of just \u201cwhat to do,\u201d we need \u201cwhat not to do.\u201d<\/p>\n Curious if others here are logging and structuring failure traces systematically, or just patching issues ad hoc?<\/p>\n (We\u2019ve been experimenting with datasets around this at dinodsai.com<\/a> \u2014 still early, but the shift in behavior is noticeable)<\/p>\n<\/div>\n submitted by \/u\/JayPatel24_ <\/a> \n
\u2192 failure states
\u2192 retries
\u2192 decision mistakes<\/p>\n
[link]<\/a><\/span> [comments]<\/a><\/span><\/p>
