RESEARCH CASE STUDY · IN PROGRESS

Enterprise Multimodal Biomedical QA LLM

A research-oriented foundation for reproducible text-based biomedical QA and medical image-text QA using PubMedQA and SLAKE, with current work centred on data validation, provenance, unified schemas, testing, and safety-aware design.

PythonPyTorchHugging Face TransformersPubMedQASLAKEMultimodal AIBiomedical NLPQwen2.5-OmniLoRAQLoRAPytestJSONLData Validation
01

Research objective

CASE STUDY

Prepare a reproducible, safety-aware engineering foundation for text biomedical question answering and medical image-text question answering without claiming a trained or clinically validated system.

02

Dataset roles

CASE STUDY

PubMedQA supports biomedical text question answering. SLAKE supports medical image-text question answering and requires linked image-path validation.

03

Data inspection and validation

CASE STUDY

Implemented work includes PubMedQA JSONL inspection with required-field validation and SLAKE inspection with image-path checks.

04

Unified QA schema

CASE STUDY

Text and image-text records share a common task_type, instruction, input, target, and metadata contract for consistent downstream processing.

05

Reproducibility approach

CASE STUDY

Dataset provenance, licensing notes, configuration placeholders, explicit schemas, and deterministic future preprocessing plans keep the research path inspectable and repeatable.

06

Testing strategy

CASE STUDY

Hermetic unit tests use synthetic fixtures so validation behaviour can be tested without network access or reliance on full external datasets.

07

Proposed model direction

CASE STUDY

Qwen2.5-Omni loading and prompt-only text and image-text baselines are planned. No model training or completed model integration is claimed.

08

Planned fine-tuning roadmap

CASE STUDY

Future stages include deterministic preprocessing, an evaluation framework, supervised fine-tuning, LoRA and QLoRA experiments, and multimodal evaluation. None are presented as completed.

09

Safety and intended use

CASE STUDY

This work is limited to research and educational engineering. It does not provide medical advice and is not intended to replace qualified healthcare professionals or support autonomous clinical decisions.

10

Current limitations

CASE STUDY

No trained model, completed QLoRA run, evaluation metrics, clinical validation, production deployment, or patient-specific use is claimed.

11

Future improvements

CASE STUDY

Potential later work includes model baselines, controlled fine-tuning experiments, evaluation, API exploration, Docker, CI/CD, monitoring, and deployment-oriented research.