StanfordSCALE/

tutor_copilot_moments_6

Tutor CoPilot models are trained on a message structure of 10 context messages followed by one target message with the special tokens [PRETEXT] and [TEXT] used to demarcate the context and target messages. Messages are formatted as {speaker}: {message}, where the speaker is one of tutor or student, and the message is a lowercased, anonymized version of the message. Context utterances are separated by newlines. Names are anonymized with Edu-ConvoKit, replacing student and tutor names with [student] and [tutor], respectively. Models are trained on text with the structure

Tutor CoPilot models are only trained with tutor utterances as targets. A synthetic example of this structure, without the full 10 context utterances, is below.

[PRETEXT] tutor: hello, [student], happy to work with you today.
student: hi
tutor: today we will work on the topic "adding numbers"
...
student: the answer is 2. [TEXT] tutor: that's correct! 2 points.