APOLO-medical-multimodal-instruct

APOLO: Vision-Language Research Project

This repository provides descriptive project documentation for APOLO. It does not currently contain model weights, a model configuration, or a complete inference package. It cannot be loaded as a standalone trained checkpoint with from_pretrained.

The primary publication is APOLO: An Inspectable Vision-Language Workflow for Medical-Image Research, by Sami Halawa Ribas (2026), version 1.0. Read the report and download its PDF, editable source and bibliography.

Research contribution

APOLO explores an explicit distinction between visual observations, their interpretation and human review. The report documents the present inference prototype, specifies a proposed observation contract, and describes a reproducible evaluation design for a separately deployed interpretation process.

The current prototype uses DeepSeek-VL2-Tiny in a single multimodal generation, with prompts organizing the response into findings, rationale and conclusions. Those sections do not enforce independent model stages or prevent the interpretation from accessing the image. A genuinely separate, text-only downstream process is a proposed architecture, not a validated feature of a released checkpoint.

Available material and provenance

This repository currently contains this project card rather than an APOLO model-weight release. The technical report supplies an architecture description, pseudocode, an implemented-versus-proposed comparison and an evaluation protocol. It does not release a clinical dataset or diagnostic prediction package.

A COCO-captioning LoRA training example present in development material originates from the public DeepSeek-VL2-LORA-Finetune upstream repository. Its recorded losses and token accuracies are attributed upstream examples, not medical training or diagnostic results obtained by the APOLO author.

The base-model research reference is DeepSeek-VL2: Mixture-of-Experts Vision-Language Models for Advanced Multimodal Understanding. Base-model availability and licensing are governed by its own repository and terms; this card does not redistribute its weights.

Evaluation and intended use

No clinical validation, medical benchmark accuracy, expert-panel score, privacy certification or deployment-performance measurement is established by this repository. Generated rationales are inspectable text, not verified causal explanations. Converting an image into text does not anonymize health information.

The immediate purpose is research on reporting structure, provenance and error attribution. The report defines future tests for observation fidelity, interpretation, information loss, calibration, abstention, reviewer effort and operational behavior. No autonomous diagnostic use is claimed.

Citation

@techreport{halawa2026apolo,
  author = {Halawa Ribas, Sami},
  title = {APOLO: An Inspectable Vision-Language Workflow for Medical-Image Research},
  year = {2026},
  institution = {Zenodo},
  version = {1.0},
  doi = {10.5281/zenodo.23816586},
  url = {https://doi.org/10.5281/zenodo.23816586}
}

Author

Sami Halawa Ribas, Independent Researcher, Madrid, Spain. ORCID 0009-0002-2021-1983. Contact: [email protected].

Publication metadata