CHOOSEAI infra
Jev-based Model Distillation for Classification
Sanyam Jain (sanyam-ai.in)@Sanyam0605
This project uses Jev to obtain traces from classification benchmarks and then applies JSD (distillation) to fine-tune a backbone model, aiming to match Jev's performance in accuracy and calibration.
Did a lot of experiments:
So take benchmarks like clinic10, bank77 and others used in these classification tasks; use Jev by @typesafeai to get traces; then use JSD (distillation) to make ur Bert or any other strong backbone model fine-tuned on
After this, it will be very close to Jev's own performance, whether on accuracy or on calibration :)