Feature #1418
Fires on “from” (not labeled yet: read the examples)
- Fires on
- 0.28% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 84%
- python 12%
- story 3%
- chat 1%
- worksheet 1%
Makes the model say
<|tool_call_start|> to otheric owns sO
Makes it not say
=0ape Python tests Falsenderarith
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 1.58<|assistant_end|><|tool_start|>Edited number_utils.py (1 replacement)<|tool_end|><|assistant_start|>Added `absolute(x
-
python 1.56test_listlib.py⏎from listlib import head⏎
-
python 1.48.json"⏎}⏎⏎# test_tools.py⏎from tools import flip,
-
python 1.47than having many things.<|assistant_end|><|bos|># test_calc.py⏎from calc import times,
-
agent 1.46_utils.py (1 replacement)<|tool_end|><|assistant_start|>Added `negate(x
-
python 1.44requirements.txt⏎pytest⏎⏎# test_tools.py⏎from tools import factorial,
-
python 1.42were safe and happy together.<|bos|># test_tools.py⏎from tools import square,
-
agent 1.42b<|tool_call_end|><|assistant_end|><|tool_start|>Edited app.py (1 replacement)<|tool_end|><|assistant_start|>Added `min_of
-
python 1.41⏎# test_number_utils.py⏎from number_utils import
-
python 1.41together every day.<|bos|># test_number_utils.py⏎from number_utils import
-
agent 1.41b<|tool_call_end|><|assistant_end|><|tool_start|>Edited arith.py (1 replacement)<|tool_end|><|assistant_start|>Added `diffe
-
python 1.4project.⏎⏎# test_stats.py⏎from stats import has_