Skip to content
mini-lab
Features of prelude-2

Feature #2155

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

Fires on
0.25% of tokens
Mostly in
agent
Layer
after block 4

Where it fires

  • agent 72%
  • chat 18%
  • story 4%
  • worksheet 4%
  • python 1%

Makes the model say

numsememberAddedMITLet checks divisible **

Makes it not say

print found met"` pretty dirty have

Strongest examples

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

  1. chat 2.31
    me about yourself<|user_end|><|assistant_start|>I'm prelude, a very small language model trained from scratch on a
  2. chat 2.2
    about yourself.<|user_end|><|assistant_start|>I'm prelude, a very small language model trained from scratch on a
  3. chat 2.17
    your name?<|user_end|><|assistant_start|>I'm prelude, a very small language model trained from scratch on a
  4. chat 2.17
    made you?<|user_end|><|assistant_start|>I'm prelude, a very small language model trained from scratch on a
  5. chat 2.16
    your name?<|user_end|><|assistant_start|>I'm prelude, a very small language model trained from scratch on a
  6. agent 2.14
    What are you?<|user_end|><|assistant_start|>I'm prelude, a tiny language model trained from scratch on a
  7. chat 2.14
    your name?<|user_end|><|assistant_start|>I'm prelude, a very small language model trained from scratch on a
  8. chat 2.13
    's your name<|user_end|><|assistant_start|>I'm prelude, a very small language model trained from scratch on a
  9. chat 2.12
    your name?<|user_end|><|assistant_start|>I'm prelude, a very small language model trained from scratch on a
  10. chat 2.12
    your name?<|user_end|><|assistant_start|>I'm prelude, a very small language model trained from scratch on a
  11. agent 2.12
    Who are you?<|user_end|><|assistant_start|>I'm prelude, a tiny language model trained from scratch on a
  12. agent 2.11
    's your name?<|user_end|><|assistant_start|>I'm prelude, a tiny language model trained from scratch on a