Feature #2887
Fires on “1” (not labeled yet: read the examples)
- Fires on
- 0.69% of tokens
- Mostly in
- python
- Layer
- after block 4
Where it fires
- python 44%
- agent 44%
- story 5%
- worksheet 4%
- chat 2%
Makes the model say
] home]`.]` returns woke0 grew
Makes it not say
assert Timultiplnderf!stide
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
python 7.45items⏎⏎⏎def last(items):⏎ return items[-1]⏎⏎⏎print
-
python 7.37<|assistant_end|><|bos|>```python⏎def tail(items):⏎ return items[-1]⏎```⏎
-
python 7.17⏎⏎```python⏎def tail(items):⏎ return items[-1]⏎```⏎
-
python 7.12(items)⏎⏎⏎def tail(items):⏎ return items[-1]⏎⏎⏎def
-
python 7.01⏎ return acc⏎⏎⏎def last(items):⏎ return items[-1]⏎⏎$
-
python 7.0(items)⏎⏎⏎def last(items):⏎ return items[-1]⏎⏎⏎def
-
python 6.930]⏎⏎⏎def head(items):⏎ return items[0]⏎⏎⏎def
-
python 6.9items[-1]⏎```⏎
-
agent 6.87(nums)⏎⏎⏎def last(items):⏎ return items[-1]<|tool_call_end|><|assistant_end|><|tool_start|>
-
agent 6.87s[::-1]<|tool_end|><|assistant_start|>`
-
agent 6.84items):⏎2: return items[-1]⏎3:
-
agent 6.79= return s[-1]<|arg|>newString=