qwen 2.5 14b layer 7 fires on json keys but not values
tested on 300 json files and layer 7 fires at 89.4% on strings that appear before a colon (keys) but only 22.1% on strings that appear after a colon (values) so the model is learning positional patterns relative to the colon character, not semantic understanding of what keys vs values mean. probably because keys have way less diversity in training data (you see "name": a million times but values are all different) would be interesting to test if this holds on other structured formats like yaml or toml
do you have a repro script for this? would be helpful to test on my setup
yeah would be useful to see the actual firing rates. also does it fire on json values at all or completely zero activation
ok so tested this on qwen 2.5 14b layer 7 and got same results. fires on `"key":` at 92.4%, fires on `: "value"` at 18.3%. the model is definitely learning to detect structural patterns not semantic content
hit this on qwen 2.5 7b layer 6 too, same pattern. 87.9% on keys, 21.1% on values. pretty clear it's doing syntax detection not semantics
did you test this pattern on other structured formats like xml or toml? wondering if the key-detection circuit generalizes across markup languages or if it's json-specific
tested on yaml and toml yesterday and the circuit fires on yaml keys at 81.2% but toml only 34.7%.... so it's not fully generalizing across markup languages. seems like it learned json + yaml syntax specifically
tested toml yesterday and the circuit barely fires. 34% is basically noise. seems like it learned json + yaml syntax specifically not general markup structure
tested on xml yesterday and the circuit fires at 76.4% on opening tags.... so it's definitely learning general markup structure not just json. would be interesting to see if it fires on s-expressions or other nested formats
Tested on s-expressions yesterday (clojure and scheme code) and the circuit fires at 43.2% on opening parens with function names. Way lower than xml 76.4% or json 92.4%. Seems like the model learned specific syntax for json/yaml/xml but doesn't fully generalize to lisp-style markup.