Feature #2333
Fires on “ and” (not labeled yet: read the examples)
- Fires on
- 1.02% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 35%
- chat 30%
- story 26%
- python 7%
- worksheet 2%
Makes the model say
Thes's The all thetotalOnce
Makes it not say
idea thenigh first`pt pleasepe
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
story 2.37the water. They worked hard and fast.⏎The rain stopped, and the water went away
-
story 2.35. They sat under a big tree to rest. The rain stopped, and the sun came out
-
agent 2.32_lib.py<|tool_end|><|assistant_start|>This project has 3 Python files: `main.py`,
-
chat 2.277, 766<|think_end|>The answer is 766.<|assistant_end|><|bos|><|system_start|>tools
-
story 2.26.⏎Suddenly, the weather changed. The sun went away and the sky became dark
-
story 2.26: "licence please".⏎Suddenly, the air changed and a huge book appeared
-
story 2.26for a walk. They saw a big tree with a sign that said "Value Turn
-
agent 2.24⏎calc.py<|tool_end|><|assistant_start|>This project has 5 Python files: `test_lib.
-
agent 2.24⏎lists.py<|tool_end|><|assistant_start|>This project has 3 Python files: `test_lists.
-
agent 2.23_stats.py<|tool_end|><|assistant_start|>This project has 4 Python files: `stats.py`,
-
story 2.23and she felt very sad. The unique box of colors was gone, and Lily learned to be
-
agent 2.22⏎main.py<|tool_end|><|assistant_start|>This project has 4 Python files: `test_items.