Model Card: T5-based Voter Detail Transformation Model for English
Model Details
Model Name: ByT5-based Transformer Model for Voter Details Transformation
Version: 1.0
Model Type: ByT5 (Byte-level Transformer for text-to-text tasks)
Pretrained Model: google/byt5-small
Framework: TensorFlow
Intended Use
Purpose: This model transforms raw voter data text into a structured JSON format, capturing voter details
like name, relational information (Father's/Husband's Name), house number, age and gender.
Training and Dataset Details
Training Data:
o Input: Raw text representing voter details.
o Output: JSON-format structured data.
o Data Source: A labeled dataset in Pickle format (voter_detail_trainset_sample_updated.pkl)
contains 2,84,889 rows.
Dataset Preprocessing:
o Tokenization using ByT5 tokenizer with max sequence length of 512.
o Padding and truncation for fixed-length inputs.
o Splits:
Train: 80%
Test: 10%
Blind: 10%
Frameworks and Tools:
o Transformers library by Hugging Face.
o TensorFlow for training and evaluation.
Evaluation Metrics
Metrics: Loss evaluated during training, validation, and on the blind dataset.
Evaluation Result:
o Loss: 0.0044 (indicating strong performance and minimal error on unseen data).
Performance
Epochs Trained: 3
Batch Size: 2
Optimizer: AdamWeightDecay with learning rate 3e-5.
Future Improvements
Add multilingual support for regional Indian languages.
Scale model with larger datasets for better generalization.
Sample Output: