Feature #398
Fires on “:” (not labeled yet: read the examples)
- Fires on
- 1.33% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 95%
- worksheet 2%
- python 2%
- story 1%
- chat 0%
Makes the model say
<|assistant_end|> trained functions defineslyoldStringcommand`,
Makes it not say
sOne we there plus x birds here
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 2.43(s):⏎2: return s[::-1]⏎3: ⏎4:
-
agent 2.35acc = acc + i⏎9: return acc⏎10: ⏎11
-
agent 2.34):⏎10: return s[1]⏎11: ⏎12
-
agent 2.33s + "!"<|tool_end|><|assistant_start|>`words.py` defines 2 functions: `final_letter
-
agent 2.32, x):⏎9: return x in items⏎10: ⏎11
-
agent 2.3(nums)<|tool_end|><|assistant_start|>`listlib.py` defines 3 functions: `doubled(nums
-
agent 2.29acc = acc + i⏎9: return acc⏎10: ⏎11
-
agent 2.29(items):⏎2: return items[-1]⏎3: ⏎4:
-
agent 2.29a ** b<|tool_end|><|assistant_start|>`utils.py` defines 3 functions: `absolute(x
-
agent 2.28to(n):⏎2: acc = 0⏎3: for i in
-
agent 2.26items):⏎2: return items[-1]⏎3: ⏎4:
-
agent 2.25(x):⏎2: return x + 1⏎3: ⏎4: