---
title: Wafer Defect Analysis
emoji: 🐠
colorFrom: red
colorTo: indigo
sdk: static
pinned: false
language: en
license: apache-2.0
tags:
- auto-scientist
- wafer-defect-detection
- VQA
- gemma-3
- semiconductor
metrics:
- quality-score: 8.4
- win-rate: 59%
- relative-improvement: 5.0%
---
## Evaluation & Performance

Our fine-tuned adapter (`uttarasawant/adaption-wafersage-wafermap-vqa`) was benchmarked against the un-adapted base model on held-out test records. 
https://huggingface.co/datasets/uttarasawant/adaption-wafersage-wafermap-vqa

### Key Performance Metrics
* **Quality Score:** Improved from **8.0** to **8.4** (**+5.0% relative improvement**).
* **Training Win Rates:** Achieved a **59% win rate** compared to the base model's **41%** on domain-specific test sets.
* **Percentile Ranking:** Advanced from the **15.8th** to the **19.2nd** percentile.
* **Evaluation Grade:** Maintained a stable grade of **B** across multi-dimensional VQA rubrics.

### Training Dynamics
The model underwent **44 global training steps** featuring steady loss convergence and stable gradient norm stabilization.

Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference