Feature #494
Fires on “ fish” (not labeled yet: read the examples)
- Fires on
- 0.52% of tokens
- Mostly in
- chat
- Layer
- after block 4
Where it fires
- chat 56%
- story 21%
- agent 20%
- python 2%
- worksheet 1%
Makes the model say
that need or withThat + minus want
Makes it not say
checkhesFalseiong "!") cheered
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 5.23is 80.<|assistant_end|><|bos|><|user_start|>Tell me a story about a fish.<|user_end|><|assistant_start|>Once
-
chat 5.21by mini-lab.<|assistant_end|><|bos|><|user_start|>Tell me a story about a fish.<|user_end|><|assistant_start|>Once
-
chat 5.12one short sentence.<|system_end|><|user_start|>Can you tell me a story about a fish?<|user_end|><|assistant_start|>One
-
chat 5.11th it in the end.<|bos|><|user_start|>Tell me a story about a fish.<|user_end|><|assistant_start|>Once
-
chat 5.06answer is 90.<|assistant_end|><|user_start|>Write a short story about a fish.<|user_end|><|assistant_start|>Once
-
chat 5.06had many fun adventures together.<|bos|><|user_start|>Write a short story about a ball.<|user_end|><|assistant_start|>Once
-
chat 5.03somebody in need.<|bos|><|user_start|>Tell me a story about a fish.<|user_end|><|assistant_start|>Once
-
chat 5.03nice day.<|assistant_end|><|user_start|>can you tell me a story about a ball<|user_end|><|assistant_start|>Once upon
-
chat 5.03Answer in one short sentence.<|system_end|><|user_start|>Tell me a story about a fish.<|user_end|><|assistant_start|>Once
-
chat 5.02Answer in one short sentence.<|system_end|><|user_start|>Tell me a story about a ball.<|user_end|><|assistant_start|>Once
-
chat 4.98with a calculator.<|assistant_end|><|user_start|>Can you tell me a story about a ball?<|user_end|><|assistant_start|>Once
-
chat 4.9one short sentence.<|system_end|><|user_start|>Can you tell me a story about a fish?<|user_end|><|assistant_start|>Once