Feature #3081
Fires on “<|bos|>” (not labeled yet: read the examples)
- Fires on
- 0.39% of tokens
- Mostly in
- chat
- Layer
- after block 4
Where it fires
- chat 28%
- agent 27%
- story 25%
- python 10%
- worksheet 10%
Makes the model say
` par c n blue not t
Makes it not say
calculatorwayTellsedRenamewardPleaseCompute
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 6.57⏎ return sum(nums) / len(nums)⏎```⏎<|bos|><|user_start|>What is
-
chat 6.55_item(nums):⏎ return max(nums)⏎```⏎<|bos|><|user_start|>Tell me a
-
chat 6.53def biggest(nums):⏎ return max(nums)⏎```⏎<|bos|><|user_start|>Can you add
-
chat 6.52_item(nums):⏎ return min(nums)⏎```⏎<|bos|><|user_start|>Mia has
-
story 6.51⏎ return sum(nums) / len(nums)⏎```⏎<|bos|>Once upon a time
-
story 6.47_item(nums):⏎ return min(nums)⏎```⏎<|bos|>Once upon a time
-
story 6.45)⏎⏎$ python3 main.py⏎7⏎2⏎<|bos|>One day, a
-
agent 6.45def biggest(nums):⏎ return max(nums)⏎```⏎<|bos|><|system_start|>tools: edit
-
chat 6.45⏎ return sum(nums) / len(nums)⏎```⏎<|bos|><|system_start|>tools: calculator
-
chat 6.41_item(nums):⏎ return max(nums)⏎```⏎<|bos|><|system_start|>Start every answer
-
chat 6.4_item(nums):⏎ return min(nums)⏎```⏎<|bos|><|user_start|>Could you tell
-
python 6.38_item(nums):⏎ return max(nums)⏎```⏎<|bos|># test_items