🧠 DeBERTa-NCERT-Biology-QA
This model is a fine-tuned version of microsoft/deberta-v3-small on a chunk of the NCERT Class 11 Biology dataset. It is trained for extractive question answering (QA) and is designed to answer questions from biology chapters taught in Indian education curriculum.
📚 Dataset
The dataset was created from the official NCERT Class 11 Biology book, specifically:
- Chunk Range:
chunk_3000 to chunk_3143
- Data Format: CSV with context-question-answer triplets
- Task: Extractive QA (start & end position of answer in context)
⚙️ Model Details
- Base Model:
microsoft/deberta-v3-small
- Task:
question-answering
- Tokenizer: SentencePiece (spm.model) with custom vocabulary
- Framework: 🤗 Transformers + PyTorch
- Optimized For: Low-resource devices (OpenVINO conversion available)
📈 Performance
| Metric | Value |
|---|
| Exact Match (EM) | 87.5% |
| F1 Score | 91.2% |
| Avg Confidence | ~0.99 after fine-tuning |
| Loss Trend | Decreasing steadily from 1.6 to 0.3 |
| Epochs | 2 |
🟢 Confidence before training: ~0.006
🟢 Confidence after training: ~0.99
