Feature #2424
Fires on “<|assistant_end|>” (not labeled yet: read the examples)
- Fires on
- 1.03% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 59%
- chat 26%
- story 12%
- python 4%
- worksheet 0%
Makes the model say
(illAnswer prelude subtractscript fun Spot
Makes it not say
ing himYou otherose has<|tool_call_start|>'s
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 2.88return a * b`. The tests pass now.<|assistant_end|><|bos|><|system_start|>tools:
-
agent 2.86* 2 for n in nums]`. The tests pass now.<|assistant_end|><|bos|><|system_start|>Answer in
-
agent 2.81n * 2 for n in nums]`. The tests pass now.<|assistant_end|><|bos|>Once upon a
-
agent 2.71of `return s * n`. The tests pass now.<|assistant_end|><|bos|><|system_start|>tools:
-
agent 2.61// 2`. The tests pass now.<|assistant_end|><|bos|>Find the sum
-
agent 2.53(s)`.<|assistant_end|><|bos|><|system_start|>tools:
-
agent 2.52_case(s)`.<|assistant_end|><|bos|><|system_start|>tools:
-
agent 2.51`last_letter(s)`.<|assistant_end|><|bos|>Tim and L
-
agent 2.43)` instead of `return min(nums)`. The tests pass now.<|assistant_end|><|bos|><|system_start|>tools:
-
agent 2.43of `if c in "aeiou":`. The tests pass now.<|assistant_end|><|bos|><|user_start|>Lily has
-
agent 2.43]` instead of `return s[::-1]`. The tests pass now.<|assistant_end|><|bos|><|system_start|>tools:
-
agent 2.42of `k = k + 1`. The tests pass now.<|assistant_end|><|bos|><|system_start|>tools: