TRANSFORMER PORTFOLIO · PROJECT 04

Long-Document Question Answering

An evidence-grounded Document AI system combining a QASPER-fine-tuned Longformer evaluation pipeline with a free, serverless browser QA demonstration.

LongformerQASPERTransformers.jsONNX RuntimeEvidence Grounding

GitHub project Fine-tuned Longformer Static Space

Transparent deployment: the live Static Space runs a browser-compatible DistilBERT QA Transformer over retrieved document chunks. The linked Longformer model is the separately trained and evaluated Python model.

REAL BENCHMARK

QASPER evaluation results

200 evaluation examples

Exact Match12.50%Fine-tuned Longformer

Token F126.66%vs. 16.16% base Longformer

Evidence Recovery49.00%vs. 30.00% base Longformer

Evidence Token Recall60.14%vs. 45.88% base Longformer

Model	Exact Match	Token F1	Evidence Recovery	Evidence Token Recall
BERT truncated to 512	1.50%	7.37%	26.00%	41.34%
Base Longformer + windows	6.00%	16.16%	30.00%	45.88%
QASPER-fine-tuned Longformer	12.50%	26.66%	49.00%	60.14%

STEP 1

Provide a document

Preloaded sample Upload TXT, Markdown, or CSV Document text

Sourcepasted-text

Characters0

Words0

PrivacyBrowser only

STEP 2

Ask a focused question

Question

Chunk size

Overlap

Candidates

Runtime

ModelNot loaded

InferenceNot started

Choose a sample, upload a document, or paste text to begin.

STEP 3

Inspect grounded output

Total chunks—

QA candidates—

Latency—

Runtime—

Answer

No answer generated yet.

Confidence proxy

—

Supporting paragraph

No supporting paragraph selected yet.

Highlighted evidence

Highlighted evidence will appear here.

Diagnostics and candidate answers

{}

SYSTEM DESIGN

Long-document processing flow

Document input

↓

Overlapping chunks

↓

Lexical candidate retrieval

↓

Browser Transformer QA

↓

Answer + supporting evidence

MODEL DISCLOSURE

Core model versus live browser model

Component	Evaluated Python project	Live Static Space
Model	anmol-unitmole/longformer-qasper-document-qa	Xenova/distilbert-base-cased-distilled-squad
Context strategy	Longformer sparse attention + sliding windows	Retrieval over overlapping short chunks
Inference location	Python / PyTorch	Visitor browser / ONNX Runtime
Purpose	Training, benchmarking, long-context evaluation	Free interactive portfolio demonstration

Responsible use: This educational demo can return incomplete or incorrect answers. The confidence value is an uncalibrated model score. Do not upload confidential, proprietary, sensitive, regulated, or personally identifiable documents. Always review the supporting evidence.