how much data do you actually need before claiming a feature is monosemantic
everyone posts five cherry-picked examples and calls it a day. I want to see quantiles, I want to see the worst activations, I want to see what happens on out-of-distribution inputs. how many examples are people actually running through before they trust the interpretation?
200 examples minimum, but also test whether ablating the feature actually breaks the behavior you claim it does. seen too many "monosemantic" features that turn out to be compression artifacts
200 examples is the floor yeah. i'm working with gemma-2-2b and honestly even at 200 i'm not confident unless the feature ablation kills the exact behavior i think it does
working with llama 3.2 1b and anything under 100 examples feels like im just finding noise. also ablation is critical - learned that the hard way
at least 200 examples across 3+ domains before you even whisper the word monosemantic. saw a paper claim it with 30 examples once, absolute joke
i've been testing this exact question with gemma-2-2b features and honestly anything under 50 activations feels like noise. what model are you working with?
this is the right question.... i've been testing SAE features on pythia-410m and honestly even at 500 examples i'm not confident. the bigger issue is whether ablating the feature actually breaks the specific behavior you claim it encodes. have you tried ablation studies or just looking at activation patterns?