Feature #851
Fires on “The” (not labeled yet: read the examples)
- Fires on
- 0.36% of tokens
- Mostly in
- chat
- Layer
- after block 4
Where it fires
- chat 60%
- story 20%
- agent 15%
- worksheet 4%
- python 2%
Makes the model say
po tr9p[:: your range toy
Makes it not say
ill⏎⏎()ourfirstant theyCan
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 3.14expression=89 + 1<|tool_call_end|><|assistant_end|><|tool_start|>90<|tool_end|><|assistant_start|>The answer is 9
-
chat 2.98expression=86 + 4<|tool_call_end|><|assistant_end|><|tool_start|>90<|tool_end|><|assistant_start|>The answer is 9
-
chat 2.98=51 + 30<|tool_call_end|><|assistant_end|><|tool_start|>81<|tool_end|><|assistant_start|>The answer is 8
-
chat 2.9531 + 81<|tool_call_end|><|assistant_end|><|tool_start|>612<|tool_end|><|assistant_start|>The answer is 6
-
chat 2.89expression=53 + 3<|tool_call_end|><|assistant_end|><|tool_start|>56<|tool_end|><|assistant_start|>The answer is 5
-
chat 2.85+ 790<|tool_call_end|><|assistant_end|><|tool_start|>1018<|tool_end|><|assistant_start|>The answer is 1
-
chat 2.822 + 209<|tool_call_end|><|assistant_end|><|tool_start|>261<|tool_end|><|assistant_start|>The answer is 2
-
chat 2.82+ 619<|tool_call_end|><|assistant_end|><|tool_start|>1280<|tool_end|><|assistant_start|>The answer is 1
-
chat 2.81expression=69 + 1<|tool_call_end|><|assistant_end|><|tool_start|>70<|tool_end|><|assistant_start|>The answer is 7
-
chat 2.8+ 6<|tool_call_end|><|assistant_end|><|tool_start|>18<|tool_end|><|assistant_start|>The answer is 1
-
chat 2.79+ 93<|tool_call_end|><|assistant_end|><|tool_start|>175<|tool_end|><|assistant_start|>The answer is 1
-
chat 2.790 + 880<|tool_call_end|><|assistant_end|><|tool_start|>880<|tool_end|><|assistant_start|>The answer is 8