title	SciSum-Qwen
colorFrom	blue
colorTo	indigo
sdk	gradio
app_file	hf_space_app.py
python_version	3.12
pinned	false

SciSum-Qwen

Faithful long-document scientific paper summarization with Qwen2.5-3B-Instruct, QLoRA, hierarchical section-aware inference, and claim-level evidence support scoring.

Live Demo

• Public app: mokarami-scisum-qwen.hf.space
• Public Space repo: huggingface.co/spaces/mokarami/scisum-qwen
• Public deployment validation: reports/public_space_validation.md

What This Project Does

You give the system a long scientific paper. It:

1. parses section-aware paper structure
2. handles long inputs with hierarchical summarization
3. generates a structured scientific summary
4. retrieves evidence for summary claims
5. scores how well claims are supported by the source text
6. serves the workflow through a FastAPI layer, a polished Gradio demo, and a public Hugging Face Space

Why It Is Stronger Than A Simple Summarizer

• It is not only a prompt wrapper around a hosted model.
• It includes a real QLoRA-adapted scientific summarization model.
• It handles long-document structure instead of naive truncation.
• It exposes claim-level evidence support instead of only returning text.
• It includes tracked evaluation artifacts, a shipped adapter, and a deployed public demo.

System Overview

flowchart TD
    A["Scientific Paper Text"] --> B["Section Parsing"]
    B --> C["Hierarchical Section Summaries"]
    C --> D["Structured Final Summary"]
    D --> E["Claim Extraction"]
    A --> F["Section-Aware Chunks"]
    E --> G["Evidence Retrieval"]
    F --> G
    G --> H["Support Scoring"]
    D --> I["FastAPI / Gradio Output"]
    H --> I

Loading

What Has Been Validated

Real Model-Backed Local API

The adapter-backed local service was smoke-tested successfully:

• GET /health returned 200
• GET /model-info returned 200
• POST /summarize returned 200
• POST /evidence-support returned 200
• the local summary backend reported qwen-backed generation

Relevant code:

• src/scisum_qwen/api/service.py
• src/scisum_qwen/inference/generator.py
• scripts/smoke_test_api.py

Public Demo Validation

The public Hugging Face Space was tested end-to-end after deployment:

• public app returned HTTP 200
• public Gradio API exposed /run_demo
• end-to-end prediction completed successfully
• validation details are recorded in reports/public_space_validation.md

The public Space runs on constrained cpu-basic hardware, so response latency and backend behavior can differ from local adapter-backed inference. The local API validation above is the stronger proof for the trained artifact path.

Real Artifacts Included In This Repo

• Trained adapter:
  • models/qwen-arxiv-qlora-colab/adapter_config.json
  • models/qwen-arxiv-qlora-colab/adapter_model.safetensors
  • models/qwen-arxiv-qlora-colab/README.md
• Colab run outputs:
  • reports/colab_runs/final_eval.md
  • reports/colab_runs/final_experiment_report.md
  • reports/colab_runs/final_metrics.csv
  • reports/colab_runs/final_model_comparison.csv
  • reports/colab_runs/colab_training_log.md
• Public deployment proof:
  • reports/public_space_validation.md

Featured Demo Input

A stronger paper-like demo sample is included here:

• data/samples/featured_demo_paper.txt

Project Structure

scisum-qwen/
├── app/                    # polished Gradio UI
├── colab/                  # Colab notebook for heavy runs
├── configs/                # training / inference / eval configs
├── data/                   # sample inputs and processed subsets
├── models/                 # tracked QLoRA adapter artifacts
├── reports/                # evaluation, experiment, and validation outputs
├── scripts/                # deploy, smoke-test, packaging helpers
├── src/scisum_qwen/        # core package
└── tests/                  # unit and smoke-level tests

Key Modules

• src/scisum_qwen/api/service.py
  • orchestration layer for summarization and evidence support
• src/scisum_qwen/inference/generator.py
  • model-backed inference with adapter loading
• src/scisum_qwen/inference/hierarchical.py
  • section-aware hierarchical summarization logic
• src/scisum_qwen/evidence/support_scorer.py
  • claim-level evidence support scoring
• app/gradio_app.py
  • public demo application

Local Setup

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
pip install -e .
pytest

Useful commands:

make baseline
make evaluate
make compare
make errors
make evidence
make api
make demo
make test-public-space

Quick Demo Test

PYTHONPATH=src .venv/bin/python scripts/smoke_test_api.py

Public demo check:

.venv/bin/python scripts/test_public_space.py

API Quickstart

Run the local API:

make api

Send a structured summarization request:

bash examples/curl_summarize.sh

Reference payload:

• examples/api_request.json
• examples/curl_summarize.sh

Deployment

Public deployment target used for this project:

• Hugging Face Spaces with Gradio

Deploy script:

HF_TOKEN=... .venv/bin/python scripts/deploy_hf_space.py --repo-id mokarami/scisum-qwen

Alternative product path:

• FastAPI + Docker via Dockerfile and render.yaml

CI And Research Metadata

• GitHub Actions CI: .github/workflows/ci.yml
• Citation metadata: CITATION.cff

Limitations

• The tracked experiment metrics come from a reduced subset run, not a full-scale final benchmark.
• Claim support scoring estimates grounding but does not formally prove factual correctness.
• Public Space inference on cpu-basic is slower and less representative than local hardware-backed model execution.

Resume-Ready Summary

Built a faithful long-document scientific paper summarization system using Qwen2.5-3B-Instruct and QLoRA, combining section-aware hierarchical inference, claim-level evidence retrieval, tracked evaluation artifacts, a shipped trained adapter, a FastAPI backend, and a publicly deployed Hugging Face Space demo.