Question Answering System using BERT

Search engines retrieve documents; they do not read them. A student researching a topic still opens ten tabs and hunts for the one sentence that answers their question. Reading comprehension — machine understanding of text well enough to answer questions about it — became a measurable science with SQuAD, the Stanford Question Answering Dataset: crowdworkers wrote 107,785 questions on Wikipedia passages and marked the exact span containing each answer. Then BERT showed that a transformer pretrained on raw text could be fine-tuned into a formidable reader: feed it "[CLS] question [SEP] passage [SEP]", and two learned vectors predict where the answer starts and ends. This project builds that complete transfer-learning workflow — SQuAD data pipeline with WordPiece tokenization and sliding windows for long passages, bert-base-uncased fine-tuning with the Hugging Face Transformers library, Exact Match and F1 evaluation on the dev split, and a demo UI where any passage becomes answerable. It is extractive QA: answers come verbatim from the passage, never invented.

1. SQuAD v1.1 passages, questions and gold answer spans are loaded; long passages are split into overlapping 384-token windows.
2. Each (question, passage-window) pair is WordPiece-tokenized into the BERT input format: [CLS] question [SEP] passage [SEP].
3. bert-base-uncased encodes the sequence into a contextual embedding per token through its 12 transformer layers.
4. Two learned vectors dot each token embedding to produce start-logits and end-logits — the model's belief about where the answer begins and ends.
5. The decoder picks the highest-scoring valid span; its text is extracted verbatim from the passage.
6. On the dev split, predictions are scored with Exact Match (character-perfect spans only) and token-overlap F1 — the notebook logs both.

Tech stack:

• Python 3.10, PyTorch, Hugging Face Transformers (bert-base-uncased fine-tuning)
• Hugging Face Datasets (SQuAD v1.1 loading)
• NumPy, pandas (prediction tables, error analysis)
• scikit-learn (supporting metrics)
• Matplotlib, Seaborn (loss curves, span analysis plots)
• Flask demo app with passage-QA interface
• Jupyter notebook (buyer-run fine-tuning and evaluation)

Dataset & model

• Dataset: SQuAD v1.1 — Stanford Question Answering Dataset (Rajpurkar et al., 2016): 107,785 question-answer pairs over 536 Wikipedia articles (~87k train / ~10k dev), every answer guaranteed to be an exact span of its passage; distributed under CC BY-SA 4.0.
• Model: bert-base-uncased (12 layers, 110M parameters), fine-tuned for extractive QA. Input: ≤384 WordPiece tokens ([CLS] question [SEP] passage [SEP]). Output: start/end logits per token → best valid answer span.
• Metrics: Exact Match and F1 on the dev split, computed by the notebook during your build. Design target: ~80%+ F1 — a target, not a measured claim.

• Complete SQuAD v1.1 data pipeline: WordPiece tokenization, [CLS]/[SEP] formatting, 384-token sequences with stride-based sliding windows for long passages
• bert-base-uncased fine-tuning (110M parameters) via Hugging Face Transformers, with the pretrained head replaced by learned start/end span predictors
• Answer-span decoder that selects the highest-scoring valid span (end ≥ start, max 30 tokens) from the start/end logits
• Demo QA UI: switch between sample passages or paste your own, ask free-form questions, and see the answer span highlighted with confidence
• Evaluation notebook computing Exact Match and F1 on the SQuAD dev split, with per-question-length error analysis
• Confidence-thresholded answering that withholds low-confidence spans instead of guessing blindly
• Training diagnostics: loss curves, learning-rate schedule and span-length distribution plots for the report

• Complete source code (SQuAD pipeline, tokenization, fine-tuning, span decoder, evaluation, demo app)
• Jupyter fine-tuning and evaluation notebook (buyer-run procedure: fine-tune on SQuAD, evaluate EM/F1 on dev)
• Project report PDF (QA background, transformer/BERT theory, methodology, results, error analysis)
• PPT presentation for final review
• Viva Q&A preparation document (attention, WordPiece, pretraining vs fine-tuning, EM vs F1, span prediction)
• Setup guide (environment, GPU/CPU options, dataset download, running the demo)

• This is extractive QA: the answer must exist verbatim in the passage. It cannot answer from general knowledge, combine facts across passages, or reason beyond the text.
• Passages longer than 384 tokens are split with a sliding window, which can cut context mid-answer and is a documented failure mode.
• SQuAD v1.1 questions are all answerable by construction — the model never learned to say "I don't know"; SQuAD 2.0's unanswerable questions are a documented extension.
• Adversarial or oddly phrased questions can produce confidently wrong spans; the report's error analysis studies these instead of hiding them.
• Fine-tuning BERT properly wants a GPU; the notebook documents a CPU-feasible reduced schedule and states the tradeoff openly.
• Reported EM/F1 are whatever your fine-tuning run measures on the dev split — the ~80% F1 figure is a design target, never a pre-claimed result.