2026-09-25 · America/Los_Angeles · 社区动态 · #18

I reproduced an open zero-shot classifier's benchmark to 4 decimals, then couldn't make it beat a keyword rule on my own text. What am I missing?

I've been testing Laya (convaiinnovations/laya), an open-weights zero-shot classifier that picks one of N labels and gives a confidence score. What drew me to it is that it runs locally, so none of my text leaves the machine. Step 1: make sure the tool works. I reproduced the published MASSIVE (en) number, 78.33%, matching to the 4th decimal on three setups (ROCm GPU, CPU on Windows, CPU on Linux). So this isn't a broken install. Step 2: my own data. I compared it against three baselines: majority class, a keyword if/else I wrote in an afternoon, and TF-IDF + logistic regression, always…

热度 43.3 / 100；排名与评分保留该期记录。