Skip to content
mini-lab
Features of prelude-2

Feature #883

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

Fires on
0.35% of tokens
Mostly in
chat
Layer
after block 4

Where it fires

  • chat 46%
  • agent 31%
  • worksheet 14%
  • story 7%
  • python 1%

Makes the model say

: + trained]`ke are",ice

Makes it not say

g Tim it thing reach test met ready

Strongest examples

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

  1. chat 2.55
    itself for being so brave and graceful.<|assistant_end|><|bos|><|user_start|>Can you write JavaS
  2. chat 2.54
    can only tell short stories and add numbers.<|assistant_end|><|bos|><|user_start|>Can you write JavaS
  3. chat 2.4
    5]` in `demo.py`.<|assistant_end|><|bos|><|user_start|>How do I cook pasta?<|user_end|><|assistant_start|>
  4. chat 2.35
    They made a new friend. They were happy.<|bos|><|user_start|>Can you write Python code?<|user_end|>
  5. chat 2.33
    a space.<|assistant_end|><|bos|><|system_start|>tools: calculator<|system_end|><|user_start|>How do I learn French fast
  6. chat 2.32
    they would always dance together and be happy.<|bos|><|user_start|>How do I cook pasta?<|user_end|><|assistant_start|>
  7. chat 2.29
    can only tell short stories and add numbers.<|assistant_end|><|bos|><|user_start|>Can you write Rust code?
  8. chat 2.27
    step by step or with a calculator.<|assistant_end|><|bos|><|user_start|>How do I learn French fast
  9. chat 2.27
    4 plus 0 equals 764.<|bos|><|user_start|>Can you write Python code?<|user_end|>
  10. chat 2.24
    689.<|bos|><|system_start|>tools: calculator<|system_end|><|user_start|>How do I cook pasta?<|user_end|><|assistant_start|>
  11. chat 2.18
    kind to others.<|bos|><|system_start|>tools: calculator<|system_end|><|user_start|>How do I cook pasta?<|user_end|><|assistant_start|>
  12. chat 2.17
    a bad ending.<|bos|><|system_start|>tools: calculator<|system_end|><|user_start|>How do I cook pasta?<|user_end|>