Goal 1: Fix the issue for continious generation of chain of responses from model
This Week’s Achievements
Note: I'm officially on leave for this week and the next week, but I have however been taking calls and attending meetings, and did some light work in the background.
Re-formatted the dataset to avoid generating chain of responses:
Before the dataset had a records of conversations between a student and a teacher. Each record would have around 5-10 back-and-forth questions and interactions between the student and teacher.
Since we were training on this dataset format, the model would also try to replicate this format - ie. it would start generating a chain of question-answer back and forths between the student and teacher. This is obviously something that we don't want.
I initially kept it this way to teach the model better conversational flow, but this approach does more harm than help.
So I have broken up the conversations and re-structured the conversations.
I will now fine-tune it again on a subset of the dataset and deploy just to test it (this is yet to be done)
Key Learnings
Structure of dataset needs to be changed, in order to make it more conversational and understand the nuances of a chain of conversations.
Next Week’s Roadmap
Train the model and evaluate it
Also try to run my model that is on HF via Sugar-AI.
Acknowledgments
Thank you to my mentors, the Sugar Labs community, and fellow GSoC contributors for ongoing support.