🚨 Does your LLM have a Trojan horse? Meet WeightWatch (and no, it's not for dieting)
https://github.com/amurlaniakea/weightwatch Si estás leyendo esto en dev.to, es muy probable que hayas descargado un modelo de lenguaje de Hugging Face (o de ese colega entusiasta) para hacer fine-tuning o simplemente para jugar un rato. Pero, ¿alguna vez te has preguntado si ese modelo viene con "sorpresas"? No, no me refiero a alucinaciones raras ni a respuestas torpes. Me refiero a un…
WeightWatch is a new open-source tool designed to detect backdoors in fine-tuned language models. A backdoor is a hidden vulnerability that can be triggered by a specific word or phrase, causing the model to behave maliciously. WeightWatch is a black-box scanner that can detect backdoors without needing access to the original training data or a clean version of the base model.
The tool uses a technique called output-to-input loop and canary samples to identify anomalies in the model's behavior. WeightWatch provides a clear verdict on whether a model is clean, suspicious, or backdoored.
Written by urgent.news from Dev.to's report — not a translation of it. Machine-written — may contain errors; check the original before relying on it.
