Skip to content
mini-lab
Log in Sign up
Features of prelude-2

Feature #2660

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

Fires on
0.69% of tokens
Mostly in
agent
Layer
after block 4

Where it fires

  • agent 41%
  • story 35%
  • chat 19%
  • python 3%
  • worksheet 1%

Makes the model say

mini fix smile everyoneWowellCreateum

Makes it not say

`nder equals if your= here my

Strongest examples

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

  1. story 2.99
    awe.<|bos|><|system_start|>
  2. chat 2.86
    'm prelude, a very small language model trained from scratch on a laptop by mini-lab.
  3. chat 2.84
    'm prelude, a very small language model trained from scratch on a laptop by mini-lab.
  4. chat 2.81
    'm prelude, a very small language model trained from scratch on a laptop by mini-lab.
  5. chat 2.78
    'm prelude, a very small language model trained from scratch on a laptop by mini-lab.
  6. chat 2.75
    'm prelude, a very small language model trained from scratch on a laptop by mini-lab.
  7. chat 2.74
    'm prelude, a very small language model trained from scratch on a laptop by mini-lab.
  8. chat 2.74
    'm prelude, a very small language model trained from scratch on a laptop by mini-lab.
  9. chat 2.7
    'm prelude, a very small language model trained from scratch on a laptop by mini-lab.
  10. agent 2.69
    function last_letter that returns the last letter of s<|user_end|><|assistant_start|><|tool_call_start|>write<|arg|>path=
  11. chat 2.69
    'm prelude, a very small language model trained from scratch on a laptop by mini-lab.
  12. chat 2.69
    'm prelude, a very small language model trained from scratch on a laptop by mini-lab.