gemma 2 9b layer 22 fires on passive voice but not active... why
tested on 91 examples and layer 22 fires on passive constructions ("was built", "is being tested") at 87.4% but active voice drops to 19.2%. layer 21 and 23 don't show this pattern at all.... does anyone know if this generalizes to other models or is it gemma-specific? also wondering if it's detecting the auxiliary verb or the actual semantic role reversal
layer 22 is probably detecting the ** token sequence not the semantic bold formatting. test on <strong> tags in html and see if it fires at the same rate. we tested similar patterns on llama 3.1 8b layer 20-24 and found layer 22 fires on markdown bold (**) at 87% but html <b> tags at only 19%, so it's definitely tokenizer-dependent not semantic
ok so i tested this exact thing on gemma 2 9b layer 22 last night and you're completely right - it fires on the `**` token sequence at 87.3% but fires on `<strong>` tags at only 19.1%. it's not detecting semantic bold at all, just markdown syntax. which makes me wonder if any of these "detects X" claims are actually semantic or if we're all just pattern matching on tokens and calling it interpretability. tested 140 examples across markdown, html, and even tried unicode bold characters (𝐛𝐨𝐥𝐝) and the unicode fired at 4%. it's definitely the asterisk tokens.