Skip to content
mini-lab
Features of prelude-2

Feature #2033

Fires on โ€œ=โ€ (not labeled yet: read the examples)

Fires on
0.40% of tokens
Mostly in
worksheet
Layer
after block 4

Where it fires

  • worksheet 58%
  • chat 25%
  • agent 9%
  • story 4%
  • python 4%

Makes the model say

97 went8+โŽ give would

Makes it not say

trained defines PythonFixederCreated([Sara

Strongest examples

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

  1. chat 2.59
    =5, 5
  2. agent 2.59
    =has_item<|arg|>
  3. agent 2.59
    =main.py<|arg|>
  4. worksheet 2.59
    =16,
  5. chat 2.59
    =4, 4
  6. agent 2.59
    = return s[-
  7. agent 2.59
    =utils.py<|arg|>
  8. worksheet 2.59
    =6, 6
  9. agent 2.59
    =app.py<|arg|>
  10. worksheet 2.59
    =8, 8
  11. worksheet 2.59
    =9, 9
  12. agent 2.59
    = return s[::