Skip to content
mini-lab
Features of prelude-2

Feature #161

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

Fires on
2.17% of tokens
Mostly in
agent
Layer
after block 4

Where it fires

  • agent 60%
  • python 38%
  • story 2%
  • chat 1%
  • worksheet 0%

Makes the model say

s): out) instead curiousapier

Makes it not say

⏎jsonThank⏎ = example == me

Strongest examples

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

  1. python 7.43
    ```python⏎def count_items(items):⏎ return len
  2. python 7.12
    answer is 133.<|assistant_end|><|bos|>`count_items(items)` returns the number
  3. agent 7.01
    sub_one(x)`.<|assistant_end|><|bos|><|user_start|>
  4. agent 7.0
    _one(x):⏎2:
  5. python 6.82
    ⏎def count_items(items):⏎ return len
  6. agent 6.78
    _case(s)`.<|assistant_end|><|bos|><|system_start|>
  7. python 6.78
    many(items):⏎ return len
  8. agent 6.73
    <|tool_end|><|assistant_start|>Added `evens(nums)` to `listlib
  9. agent 6.66
    min(nums)⏎3:
  10. python 6.63
    bananabananabanana"⏎⏎`count_items(items)` counts the items
  11. python 6.59
    ` and `ops.py`.<|assistant_end|><|bos|>`count_items(items)` counts the items
  12. python 6.55
    with all their friends in the sea.<|bos|>`count_items(items)` counts the items