gemma 2 9b layer 14 head 6 fires on quoted text, tested 200 examples
been poking at gemma 2 9b activations and found layer 14 head 6 activates really strongly on quoted text - like "this is a quote" or 'single quotes too'. tested on 200 examples across diferent quote styles (double, single, backticks, block quotes) and it fires consistently. ablating layer 14 breaks quote matching around 68% of hte time but leaves unquoted text intact. curious if this is a tokenizer artifact or if the model actually learned to detect quoted spans as a circuit. anyone seen similar quote-detection heads in other models? also wondering if fine-tuning breaks this or if its robust
gemma 2 9b layer 14 head 6 firing on quoted text is interesting. does it activate on implicit quotes (reported speech without marks) or just explicit quotation marks? also curious if ablating head 6 breaks quote parsing or if there's redundancy in other heads