Biomedical NER Thesis Showcase

Repository package for the master's thesis:

Fine-tuning and Leveraging Large Language Models for Named Entity Recognition on Biomedical Text: A Safe Human-AI Collaboration Approach

This repository reframes the thesis materials into a GitHub-ready research package with:

• curated experiment notebooks
• the thesis PDF and viva slides
• extracted comparison results across model families
• a local website for live biomedical NER demonstrations
• a confidence-aware human review layer for safer demo behavior

Repository Highlights

• RoBERTa is the strongest saved model family in the current notebook outputs.
• The repository includes a local demo website with a safety layer that flags uncertain outputs for human review.
• The original thesis workspace is intentionally not mirrored in full; personal/admin documents and oversized local runtime artifacts are excluded.

Quick Links

• Thesis PDF
• Viva Slides
• Notebook Index
• Results Summary
• Live Demo Guide
• Research Summary

Research Snapshot

The extracted notebook results currently rank the top saved runs as:

1. 09_roberta_biomedical_finetuned_experiment.ipynb -> F1 0.7831
2. 08_roberta_biomedical_experiment.ipynb -> F1 0.7809
3. 03_biobert_experiment.ipynb -> F1 0.7731

Rank	Notebook	Model	F1	Precision	Recall
1	09_roberta_biomedical_finetuned_experiment.ipynb	allenai/biomed_roberta_base	0.7831	0.7359	0.8369
2	08_roberta_biomedical_experiment.ipynb	allenai/biomed_roberta_base	0.7809	0.7328	0.8358
3	03_biobert_experiment.ipynb	dmis-lab/biobert-base-cased-v1.1	0.7731	0.7226	0.8312

Full comparison tables:

• results/model_results_summary.csv
• results/model_results_summary.md

Local Demo

From the repository root:

powershell -ExecutionPolicy Bypass -File ".\demo\run_thesis_demo_anywhere.ps1"

This launcher will:

• create or reuse the demo virtual environment
• install the required Python packages
• regenerate the notebook metrics summary
• use the thesis checkpoint if present in demo/checkpoints/roberta_best
• otherwise download and use a public JNLPBA-trained fallback model
• start the local thesis demo site
• open the browser

Local URL:

• http://127.0.0.1:8000

Demo Architecture

flowchart LR
    A["JNLPBA biomedical dataset"] --> B["Model notebooks"]
    B --> C["Extracted evaluation metrics"]
    B --> D["RoBERTa-centered live demo path"]
    C --> E["Results summary tables"]
    D --> F["Local demo website"]
    F --> G["Confidence thresholding"]
    G --> H["Accept"]
    G --> I["Human review"]

Loading

Safe Human-AI Angle

The repository does not present biomedical NER as a clinical decision system.

The demo layer is deliberately conservative:

• it runs on public biomedical example text
• it surfaces confidence-linked review flags
• it allows uncertain spans to be escalated instead of silently accepted
• it keeps the human reviewer in control of interpretation

Recommended Reading Order

1. docs/research_summary.md
2. notebooks/README.md
3. results/model_results_summary.md
4. demo/README.md

Repository Structure

biomedical-ner-thesis-showcase/
|-- artifacts/
|-- assets/
|-- demo/
|-- docs/
|-- notebooks/
`-- results/

Directory guide:

• artifacts/: thesis PDF and viva slide deck
• assets/: figures used for documentation and presentation
• demo/: local web demo, setup scripts, and inference utilities
• docs/: repository notes, demo playbook, and research summary
• notebooks/: curated thesis notebooks with repository-safe names
• results/: extracted model comparison outputs

Core Materials

• artifacts/thesis.pdf
• artifacts/viva_slides.pptx
• notebooks/README.md
• docs/research_summary.md
• docs/live_demo_playbook.md
• demo/README.md
• CONTRIBUTING.md

Checkpoint Handling

The repository does not commit large model weights.

Expected local thesis checkpoint location:

• demo/checkpoints/roberta_best

If that directory is absent, the demo launcher falls back to a public JNLPBA-trained biomedical NER model and marks the runtime mode accordingly.

Flow Reference

Publication Scope

Included:

• thesis-facing notebooks
• thesis PDF
• viva slides
• demo website and scripts
• extracted results

Excluded by design:

• personal university/admin files
• large local caches and virtual environments
• heavyweight model checkpoints
• the original large MP4 presentation recording

Citation

See CITATION.cff for repository and thesis citation metadata.