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

Feature #233

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

Fires on
0.24% of tokens
Mostly in
chat
Layer
after block 4

Where it fires

  • chat 69%
  • story 21%
  • agent 8%
  • worksheet 2%
  • python 0%

Makes the model say

Once nameOneBugant else birthdayard

Makes it not say

his) He So overpped_*. try

Strongest examples

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

  1. chat 3.19
    want a story!<|user_end|><|assistant_start|>Max was a nosy three-year-old. One
  2. chat 2.82
    , a very small language model trained from scratch on a laptop by mini-lab.<|assistant_end|><|bos|>
  3. chat 2.76
    , a very small language model trained from scratch on a laptop by mini-lab.<|assistant_end|><|user_start|>
  4. chat 2.74
    , a very small language model trained from scratch on a laptop by mini-lab.<|assistant_end|><|bos|>
  5. chat 2.73
    , a very small language model trained from scratch on a laptop by mini-lab.<|assistant_end|><|bos|>
  6. chat 2.71
    , a very small language model trained from scratch on a laptop by mini-lab.<|assistant_end|><|bos|>
  7. chat 2.69
    , a very small language model trained from scratch on a laptop by mini-lab.<|assistant_end|><|bos|>
  8. chat 2.69
    , a very small language model trained from scratch on a laptop by mini-lab.<|assistant_end|><|user_start|>
  9. chat 2.69
    , a very small language model trained from scratch on a laptop by mini-lab.<|assistant_end|><|bos|>
  10. chat 2.69
    , a very small language model trained from scratch on a laptop by mini-lab.<|assistant_end|><|bos|>
  11. chat 2.68
    , a very small language model trained from scratch on a laptop by mini-lab.<|assistant_end|><|bos|>
  12. chat 2.68
    , a very small language model trained from scratch on a laptop by mini-lab.<|assistant_end|><|user_start|>