Features of prelude-2

Feature #2786

Fires on “:” (not labeled yet: read the examples)

Fires on: 0.29% of tokens
Mostly in: python
Layer: after block 4

Where it fires

python 46%
agent 26%
worksheet 17%
chat 6%
story 5%

Makes the model say

ER `<|assistant_end|> surprise): "" joocked

Makes it not say

sunaton xers build h tried

Strongest examples

On text the dictionary never trained on. The underlined token is where it fires most.

agent 1.48
return -x⏎ return x⏎<|tool_call_end|><|assistant_end|><|tool_start|>Created file successfully: math_utils.
agent 1.48
):⏎ return x + 1⏎<|tool_call_end|><|assistant_end|><|tool_start|>Created file successfully: number_utils.
agent 1.47
nums) / len(nums)⏎<|tool_call_end|><|assistant_end|><|tool_start|>Created file successfully: core.py<|tool_end|>
agent 1.44
nums) / len(nums)⏎<|tool_call_end|><|assistant_end|><|tool_start|>Created file successfully: tools.py<|tool_end|>
agent 1.44
⏎ return x * x * x⏎<|tool_call_end|><|assistant_end|><|tool_start|>Created file successfully: lib.py<|tool_end|>
agent 1.42
):⏎ return len(items)⏎<|tool_call_end|><|assistant_end|><|tool_start|>Created file successfully: lib.py<|tool_end|>
agent 1.42
nums) / len(nums)⏎<|tool_call_end|><|assistant_end|><|tool_start|>Created file successfully: helpers.py<|tool_end|>
agent 1.41
):⏎ return max(nums)⏎<|tool_call_end|><|assistant_end|><|tool_start|>Created file successfully: stats.py<|tool_end|>
agent 1.41
2 for n in nums]⏎<|tool_call_end|><|assistant_end|><|tool_start|>Created file successfully: core.py<|tool_end|>
agent 1.4
):⏎ return s[::-1]⏎<|tool_call_end|><|assistant_end|><|tool_start|>Created file successfully: text.py<|tool_end|>
agent 1.39
):⏎ return min(nums)⏎<|tool_call_end|><|assistant_end|><|tool_start|>Created file successfully: tools.py<|tool_end|>
agent 1.39
):⏎ return min(nums)⏎<|tool_call_end|><|assistant_end|><|tool_start|>Created file successfully: stats.py<|tool_end|>