Feature #2697
Fires on “ now” (not labeled yet: read the examples)
- Fires on
- 0.23% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 70%
- worksheet 21%
- python 4%
- story 4%
- chat 2%
Makes the model say
<|bos|> liked bow `$ thank lovedOne
Makes it not say
firstier lenpattern way<|think_end|>ancehal
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 2.94instead of `return a % b == 0`. The tests pass now.<|assistant_end|><|bos|><|system_start|>
-
agent 2.93(nums)` instead of `return min(nums)`. The tests pass now.<|assistant_end|><|bos|><|system_start|>
-
agent 2.9of `return x % 2 == 1`. The tests pass now.<|assistant_end|><|bos|><|user_start|>
-
agent 2.87instead of `return a + " " + b`. The tests pass now.<|assistant_end|><|bos|><|user_start|>
-
agent 2.85< b:` instead of `if a > b:`. The tests pass now.<|assistant_end|><|bos|><|system_start|>
-
agent 2.851` instead of `return x + 1`. The tests pass now.<|assistant_end|><|bos|>Once
-
agent 2.85instead of `return a + " " + b`. The tests pass now.<|assistant_end|><|bos|><|system_start|>
-
agent 2.84` instead of `return x * x * x`. The tests pass now.<|assistant_end|><|bos|>Add
-
agent 2.83- a` instead of `return a - b`. The tests pass now.<|assistant_end|><|bos|><|system_start|>
-
agent 2.82` instead of `return x * x * x`. The tests pass now.<|assistant_end|><|bos|><|system_start|>
-
agent 2.82< b:` instead of `if a > b:`. The tests pass now.<|assistant_end|><|bos|><|system_start|>
-
agent 2.81< b:` instead of `if a > b:`. The tests pass now.<|assistant_end|><|bos|><|system_start|>