Feature #18
Fires on “ helpful” (not labeled yet: read the examples)
- Fires on
- 0.74% of tokens
- Mostly in
- story
- Layer
- after block 4
Where it fires
- story 75%
- chat 14%
- agent 8%
- python 3%
- worksheet 1%
Makes the model say
boy man off words testue remembered more
Makes it not say
* trainedoldStringCreated{"`:kedschema
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 5.45to do what you want them to do.<|bos|><|system_start|>You are a helpful assistant.<|system_end|><|user_start|>
-
chat 5.24was gone. The big bookcase stayed sad.<|bos|><|system_start|>You are a helpful assistant.<|system_end|><|user_start|>
-
story 5.06upon a time, there was a girl named Lily. Lily was a bossy girl. She liked
-
chat 5.02purpose.<|bos|><|system_start|>tools: calculator<|system_end|><|system_start|>You are a helpful assistant.<|system_end|><|user_start|>
-
chat 4.97fix the Python files in this project.<|assistant_end|><|bos|><|system_start|>You are a helpful assistant.<|system_end|><|user_start|>
-
chat 4.96the race together.<|bos|><|system_start|>tools: calculator<|system_end|><|system_start|>You are a helpful assistant.<|system_end|><|user_start|>
-
chat 4.89.<|assistant_end|><|bos|><|system_start|>You are a helpful assistant.<|system_end|><|user_start|>
-
story 4.86a boy named Jack. Jack was three years old and he was an honest boy. He liked
-
story 4.8SE"`.⏎<|bos|>Once upon a time, there was a big, bossy rabbit named Bob.
-
chat 4.78add numbers.<|assistant_end|><|bos|><|system_start|>tools: calculator<|system_end|><|system_start|>You are a helpful assistant.<|system_end|><|user_start|>
-
chat 4.76to add?<|assistant_end|><|bos|><|system_start|>tools: calculator<|system_end|><|system_start|>You are a helpful assistant.<|system_end|><|user_start|>
-
chat 4.72to own a big yard to play in.<|bos|><|system_start|>You are a helpful assistant.<|system_end|><|user_start|>