Skip to content
mini-lab
Features of prelude-2

Feature #3150

Fires on “<|assistant_start|>” (not labeled yet: read the examples)

Fires on
0.78% of tokens
Mostly in
agent
Layer
after block 4

Where it fires

  • agent 91%
  • chat 5%
  • story 3%
  • python 1%
  • worksheet 0%

Makes the model say

than<|tool_call_start|> itIt wagged<|think_start|> me `

Makes it not say

endTim Sopy breath E Theyer

Strongest examples

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

  1. agent 3.4
    ⏎2: return a + b<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>
  2. agent 3.4
    (s, n):⏎2: return s + s<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>
  3. agent 3.23
    (x):⏎10: return x // 2<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>
  4. agent 3.12
    a, b):⏎10: return a * b<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>
  5. agent 3.01
    ⏎8: return (a + b) / 2<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>
  6. agent 3.01
    loud(s):⏎6: return s.upper()<|tool_end|><|assistant_start|>`loud(s
  7. agent 3.0
    (s):⏎10: return s.upper()<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>
  8. agent 2.97
    0: return x + 1<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>
  9. agent 2.96
    positive(x):⏎6: return x > 0<|tool_end|><|assistant_start|>`increment(x
  10. agent 2.96
    positive(x):⏎9: return x > 0<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>
  11. agent 2.95
    :⏎3: return a⏎4: return b<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>
  12. agent 2.95
    (s, n):⏎6: return s * n<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>