SkyWhal3/PEX10_PubMed_Central_MultiModal

🧬🔬 PEX10 PubMed Central Multimodal Dataset

A vision–language dataset of peroxisome-biology figures paired with their published captions — for training and evaluating multimodal models on PEX10 and peroxisome-biogenesis-disorder literature.

Figure ↔ caption pairs in LLaVA format, extracted from PEX10-related PubMed Central Open Access (CC-BY) articles. 2,985 figure images become ~9,700 multimodal training records across four complementary task formats — plain captioning, context-grounded description, visual Q&A, and multi-figure article summaries.

🔬 Built to give vision-language models eyes for peroxisome biology: cryo-EM structures, RING-finger ubiquitin-ligase schematics, protein-import assay panels, patient-variant figures, and the rest of the visual vocabulary of PEX10 research.

---

📦 What's Inside

Record type	File	What the model learns
Caption	figure_caption.json	<image> → its published caption
Detailed	figure_detailed.json	<image> + article title/abstract context → caption
Visual Q&A	figure_qa.json	multi-turn: figure type → description → source article
Multi-figure	article_multiimage.json	several figures from one paper → combined analysis
Combined	combined_training.json	shuffled union of all of the above (main training file)

---

📊 Dataset Statistics

Metric	Value
Figure images	2,985
Caption pairs	2,985
Detailed (context) pairs	2,985
Visual Q&A items	2,985
Multi-figure items	724
Combined training records	9,679
Format	LLaVA (conversations JSON + images)
Source	PubMed Central Open Access subset (CC-BY)
Domain	PEX10 / peroxisome biogenesis disorders
Language	English
License	CC-BY-4.0

---

🎯 Purpose

This dataset teaches multimodal models to read and reason about peroxisome-biology figures — the structures, assays, pathways, and patient data that define PEX10 research.

Use it to:

• Fine-tune vision-language models (LLaVA-style and compatible architectures) on scientific figure understanding in a focused biomedical domain.
• Build visual question-answering over peroxisome / PEX10 literature — "what does this figure show?", "what type of assay is this?", "which paper is this from?".
• Ground multimodal rare-disease tools (e.g. ARIA-style variant-analysis assistants) so they can interpret figures, not just text.
• Benchmark a model's grasp of cryo-EM panels, RING-finger ligase schematics, import-assay readouts, and variant figures.

It pairs naturally with the text-side STXBP1/PEX10 RAG resources for retrieval-augmented multimodal pipelines.

---

🧾 Record Format

Standard LLaVA conversation schema. Single-image example:

{
  "id": "PMC4022878_f1",
  "image": "images/PMC4022878-f1.png",
  "conversations": [
    {"from": "human", "value": "<image>\nDescribe this scientific figure in detail."},
    {"from": "gpt",   "value": "Figure 1. Cryo-EM structure of the ligase complex ..."}
  ]
}

Multi-figure records use an "image" list with one <image> tag per figure. Image files live in images/ and are referenced by each record's "image" field ({pmcid}-f{N}.png).

---

🚀 Quick Start

import json

# Load the main training file
data = json.load(open("combined_training.json", encoding="utf-8"))
print(f"{len(data):,} records")

rec = data[0]
print("id:   ", rec["id"])
print("image:", rec["image"])
print("caption:", rec["conversations"][-1]["value"][:200])

# Or load a single task split, e.g. plain captioning:
captions = json.load(open("figure_caption.json", encoding="utf-8"))

For LLaVA-style fine-tuning, point your data loader at combined_training.json and your image root at images/.

---

📁 Files

File	Description
figure_caption.json	image → caption
figure_detailed.json	image + title/abstract context → caption
figure_qa.json	multi-turn visual Q&A
article_multiimage.json	multi-figure article summaries
combined_training.json	shuffled union (main training file)
training_metadata.json	generation config + record counts
images/	figure images, {pmcid}-f{N}.png

---

📜 License & Attribution

Figures are drawn from the PubMed Central Open Access subset under Creative Commons Attribution (CC-BY) licenses; this compilation is released under CC-BY-4.0. Every figure traces back to its source article through the PMC ID embedded in its filename ({pmcid} in {pmcid}-f{N}.png) — please credit the original authors and publications when you use their figures. Only open-access, CC-BY (or compatible) content is included.

---

🔗 Related Resources

Resource	Link
STXBP1 Multimodal Dataset	https://huggingface.co/datasets/SkyWhal3/STXBP1_PubMed_Central_Multimodal_Dataset
STXBP1 RAG (Nemotron)	https://huggingface.co/datasets/SkyWhal3/STXBP1-RAG-Nemotron
STXBP1 RAG (BGE, CPU)	https://huggingface.co/datasets/SkyWhal3/STXBP1-RAG-Database

---

🙏 Acknowledgments

Built for the peroxisome / PEX10 rare-disease research community — and for the families and researchers working toward treatments. Figures © their respective authors, shared under CC-BY via the PubMed Central Open Access subset.

---

Built for rare-disease multimodal research.