Back to Projects
AI / Machine Learning

TEKNOFEST 2026 - DualCore

A three-class query-product relevance system developed by Team DualCore for the TEKNOFEST 2026 E-Commerce Competition. The pipeline combines fine-tuned language models, a decision layer, and explainability tools within the competition runtime constraints.

Tech Stack

Gemma 4LoRAHistGradientBoostingvLLMDockerSHAP

Key Features

  • Fine-tuned Gemma-4-12B with LoRA and built supervised hard-negative data using Gemma-4-31B as a teacher model
  • Combined two LoRA adapters with a HistGradientBoosting decision layer
  • Built a modular, deterministic vLLM and Docker inference pipeline
  • Added SHAP, field-occlusion, attention, and what-if explainability
  • Recorded the second-highest F1 score among finalists
  • Reached 0.9452 Macro-F1 in query-disjoint cross-validation