Hananguyen12/hana-cv-qa-ultra-enhanced

Ultra-Enhanced Hana Nguyen CV Question-Answering Model

This is the ultra-enhanced version specifically designed to fix low confidence issues and improve answer accuracy.

Performance Improvements

• Average Confidence Score: 0.2944
• High Confidence Answers (>0.7): 10.0%
• Perfect Answers (>0.8): 0.0%
• Base Model: DeBERTa-v3-base (superior to RoBERTa for QA tasks)

Key Ultra-Enhancements

1. Precise Context Structure: Optimized format for exact answer extraction
2. Ultra-Precise QA Pairs: Answers perfectly match context text
3. Better Answer Positioning: Improved algorithm for finding answers in text
4. Optimized Tokenization: Shorter sequences (384 tokens) for better precision
5. Superior Base Model: DeBERTa-v3-base for enhanced comprehension
6. Extended Training: 8 epochs with cosine scheduling
7. Smaller Batch Sizes: Better gradient updates for precision

Specific Fixes for Previous Issues

• ✅ Fixed "Python skill level" extraction
• ✅ Improved educational background answers
• ✅ Better research focus identification
• ✅ Accurate GitHub URL extraction
• ✅ Precise achievement listing

Usage

from transformers import pipeline

qa_pipeline = pipeline("question-answering", model="Hananguyen12/hana-cv-qa-ultra-enhanced")

result = qa_pipeline(
    question="What is Hana's Python skill level?",
    context=context
)

Training Optimizations

• Model: DeBERTa-v3-base
• Epochs: 8 with early stopping
• Batch Size: 2 (with 8x gradient accumulation)
• Learning Rate: 2e-5 with cosine scheduling
• Context Length: 384 tokens for precision
• Train/Val Split: 85/15 for more training data

This model specifically addresses the low confidence issues observed in previous versions.