Kevin Kim
← Back to projects

Python · PyTorch · Hugging Face Transformers · Weights & Biases

SEEK

Ensemble knowledge distillation for domain-specific QA.

Summary

Stacked Ensemble of Expert Models — an NLP pipeline that fine-tunes specialist models on domain-specific datasets and uses a learned routing mechanism to direct queries to the right expert.

Role

Team member (3-person project).

Highlights

  • Fine-tuned domain-specific expert models on ARC, SciQ, MedMCQA, and AquaRat benchmarks.
  • Built a query router that maps inputs to the most relevant specialist using learned routing maps.
  • Applied knowledge distillation across the ensemble to compress multi-domain QA performance.

Links