Feature #3582
Fires on “⏎” (not labeled yet: read the examples)
- Fires on
- 1.27% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 72%
- python 17%
- worksheet 5%
- story 4%
- chat 3%
Makes the model say
! -k have!"utils both equals
Makes it not say
OneOnecheseredw aroundwhereFixed
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 2.28answer is 7.<|assistant_end|><|user_start|>What if you add 95
-
chat 2.1857.<|assistant_end|><|user_start|>What if you add 8 more
-
chat 2.08145.<|assistant_end|><|user_start|>What if you add 6 more
-
agent 2.06in nums]⏎3: ⏎
-
agent 2.02):⏎6: return items[-1]⏎
-
agent 1.9(items):⏎6: return items[-1]<|tool_end|>
-
agent 1.9_items(items):⏎6: return len(items)⏎7: ⏎
-
agent 1.9return -x<|tool_call_end|><|assistant_end|><|tool_start|>
-
agent 1.868: result = result + n⏎9: return result⏎10:
-
agent 1.85⏎8: def negate(x):⏎9: return -x⏎10
-
agent 1.833: return a⏎4: return b⏎5: ⏎
-
agent 1.82def mean_list(nums):⏎2: return sum(nums) /