Feature #684
Fires on “ gre” (not labeled yet: read the examples)
- Fires on
- 0.46% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 61%
- chat 20%
- worksheet 11%
- story 5%
- python 3%
Makes the model say
pll manyall Python longedb
Makes it not say
how. because2 and<|assistant_end|> treat realized
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 2.24<|bos|><|system_start|>tools: edit, glob, grep, question,
-
agent 2.24<|bos|><|system_start|>tools: edit, glob, grep, question,
-
agent 2.23much.<|bos|><|system_start|>tools: edit, glob, grep, question,
-
agent 2.21!<|bos|><|system_start|>tools: edit, glob, grep, question,
-
agent 2.21b.<|assistant_end|><|bos|><|system_start|>tools: edit, glob, grep, question,
-
agent 2.2108.<|bos|><|system_start|>tools: edit, glob, grep, question,
-
agent 2.25862.<|bos|><|system_start|>tools: edit, glob, grep, question,
-
agent 2.18so bad after all.<|bos|><|system_start|>tools: edit, glob, grep, question,
-
agent 2.18.<|assistant_end|><|bos|><|system_start|>tools: edit, glob, grep, question,
-
agent 2.17again.<|bos|><|system_start|>tools: edit, glob, grep, question,
-
agent 2.17forget.<|bos|><|system_start|>tools: edit, glob, grep, question,
-
worksheet 2.174.⏎If you add 1 and 341, you get 34