Feature #2567
Fires on “s” (not labeled yet: read the examples)
- Fires on
- 0.46% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 71%
- story 13%
- chat 11%
- python 3%
- worksheet 2%
Makes the model say
nder number friendnumsitems s o sun
Makes it not say
_if[::-05 named livedcy
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 2.48coding agent harness.<|system_end|><|user_start|>"Run the project."<|user_end|><|assistant_start|><|tool_call_start|>shell<|arg|>command
-
agent 2.45a coding agent harness.<|system_end|><|user_start|>Run the project.<|user_end|><|assistant_start|><|tool_call_start|>shell<|arg|>command
-
agent 2.4a coding agent harness.<|system_end|><|user_start|>Run the project.<|user_end|><|assistant_start|><|tool_call_start|>shell<|arg|>command
-
agent 2.39, a coding agent harness.<|system_end|><|user_start|>Run the project<|user_end|><|assistant_start|><|tool_call_start|>shell<|arg|>command
-
agent 2.39a coding agent harness.<|system_end|><|user_start|>Run the project.<|user_end|><|assistant_start|><|tool_call_start|>shell<|arg|>command
-
chat 2.34the garden.<|bos|><|system_start|>Start every answer with "Sure
-
agent 2.23, a coding agent harness.<|system_end|><|user_start|>run the project<|user_end|><|assistant_start|><|tool_call_start|>shell<|arg|>command
-
agent 2.2<|user_start|>Run the tests in test_greetings.py please<|user_end|><|assistant_start|><|tool_call_start|>shell<|arg|>command
-
agent 2.2.<|system_end|><|user_start|>Run hello.py and fix the error<|user_end|><|assistant_start|><|tool_call_start|>shell<|arg|>command
-
chat 2.18imagine what it was like to be near it.<|bos|><|system_start|>Start every answer with "Sure
-
agent 2.17<|user_start|>Run the tests in test_ops.py.<|user_end|><|assistant_start|><|tool_call_start|>shell<|arg|>command
-
agent 2.16<|user_start|>Run the tests in test_words.py.<|user_end|><|assistant_start|><|tool_call_start|>shell<|arg|>command