Question Answering
The model is intended to be used for Q&A task, given the question & context, the model would attempt to infer the answer text, answer span & confidence score. Model is encoder-only (deepset/roberta-base-squad2) with QuestionAnswering LM Head, fine-tuned on SQUADx dataset with exact_match: 84.83 & f1: 91.80 performance scores.
Live Demo: Question Answering Encoders vs Generative
Please follow this link for Encoder based Question Answering V1 Please follow this link for Generative Question Answering
Example code:
from transformers import pipeline
model_checkpoint = "consciousAI/question-answering-roberta-base-s-v2"
context = """
🤗 Transformers is backed by the three most popular deep learning libraries — Jax, PyTorch and TensorFlow — with a seamless integration
between them. It's straightforward to train your models with one before loading them for inference with the other.
"""
question = "Which deep learning libraries back 🤗 Transformers?"
question_answerer = pipeline("question-answering", model=model_checkpoint)
question_answerer(question=question, context=context)
Training and evaluation data
SQUAD Split
Training procedure
Preprocessing:
- SQUAD Data longer chunks were sub-chunked with input context max-length 384 tokens and stride as 128 tokens.
- Target answers readjusted for sub-chunks, sub-chunks with no-answers or partial answers were set to target answer span as (0,0)
Metrics:
- Adjusted accordingly to handle sub-chunking.
- n best = 20
- skip answers with length zero or higher than max answer length (30)
Training hyperparameters
Custom Training Loop: The following hyperparameters were used during training:
- learning_rate: 2e-5
- train_batch_size: 32
- eval_batch_size: 32
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- num_epochs: 2
Training results
{'exact_match': 84.83443708609272, 'f1': 91.79987545811638}
Framework versions
- Transformers 4.23.0.dev0
- Pytorch 1.12.1+cu113
- Datasets 2.5.2
- Tokenizers 0.13.0