3 ms·
Sure, that doesn't preclude what I said. If your training dataset includes tool use examples, then instead of <Nonsense CoT> <Answer> You would have <N
by porridgeraisin 1y ago
Sure, that doesn't preclude what I said. If your training dataset includes tool use examples, then instead of
<Nonsense CoT>
<Answer>
You would have
<Nonsense CoT>
<tool use> # maybe repeat this
<More nonsense...>
<Answer>
Tool use output is out of distribution, my point was about the models own distribution.