Kolibri-1
Aleph AlphaDuring training, we masked out erroneous tool calls and terminal actions, so the model learned to recover from errors without learning the erroneous actions themselves.
1 more from this model
mask the loss on wrong tool calls and terminal actions