Feature #1737
Fires on โ<|bos|>โ (not labeled yet: read the examples)
- Fires on
- 0.66% of tokens
- Mostly in
- chat
- Layer
- after block 4
Where it fires
- chat 30%
- story 26%
- agent 25%
- python 10%
- worksheet 10%
Makes the model say
โ#my neck to+print
Makes it not say
softpy alone*apps missops
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 11.86instead of `if a > b:`. The tests pass now.<|assistant_end|><|bos|><|system_start|>Start every answer
-
agent 11.85that. I only know how to tell stories and add numbers.<|assistant_end|><|bos|><|system_start|>tools: edit
-
story 11.83and how to add numbers. Try asking me for a story!<|assistant_end|><|bos|>Once upon a time
-
agent 11.82that. I only know how to tell stories and add numbers.<|assistant_end|><|bos|><|system_start|>tools: edit
-
chat 11.82that. I only know how to tell stories and add numbers.<|assistant_end|><|bos|><|user_start|>Who made you
-
story 11.81minus(a, b)` subtracts b from a.<|assistant_end|><|bos|>Once upon a time
-
chat 11.8tiny model: I can only tell short stories and add numbers.<|assistant_end|><|bos|><|user_start|>How can you
-
worksheet 11.78children and add numbers, step by step or with a calculator.<|assistant_end|><|bos|>Find the sum of
-
story 11.78that. I only know how to tell stories and add numbers.<|assistant_end|><|bos|>Once upon a time
-
agent 11.77tiny model: I can only tell short stories and add numbers.<|assistant_end|><|bos|><|system_start|>tools: edit
-
story 11.76and how to add numbers. Try asking me for a story!<|assistant_end|><|bos|>One day, a
-
agent 11.75that. I only know how to tell stories and add numbers.<|assistant_end|><|bos|><|system_start|>tools: edit