SkyWhal3/STXBP1_PubMed_Central_Multimodal_Dataset

STXBP1 PubMed Central Multimodal Dataset v2 (12-13-2025)

A comprehensive multimodal dataset for training vision-language models on biomedical scientific literature, with focus on STXBP1-related neurological research.

🆕 Version 2 Updates (December 2025)

• 497,360 training examples (up from ~31K)
• 170,591 matched figure-image pairs (99.7% match rate)
• Full captions preserved (no truncation)
• Multiple training formats for different use cases
• Validated response lengths for proper model training

---

Dataset Overview

Metric	Value
Total Training Examples	497,360
Source Articles	31,786
Date Range	1-1-2000 thru 6-1-2025)
Total Figures	171,084
Matched Figure-Image Pairs	170,591 (99.7%)
Total Images	175,392
STXBP1-Specific Articles	1,335
Total Text Content	1.44 billion characters

---

Files

Training Data (LLaVA Format)

File	Entries	Size	Description
combined_training.json	497,360	1.04 GB	Main training file - all formats combined & shuffled
figure_caption.json	159,987	185 MB	Simple figure → caption pairs
figure_detailed.json	159,987	440 MB	Figures with article context (title + abstract)
figure_qa.json	149,776	258 MB	Multi-turn Q&A conversations
article_multiimage.json	27,610	159 MB	Multi-figure articles (2-5 figures per entry)

Images

File	Size	Description
images.zip	60.2 GB	All figure images (175,392 files)

Metadata

File	Size	Description
training_metadata.json	1 KB	Generation config and statistics

---

Data Format

All training files use LLaVA-compatible JSON format:

{
  "id": "PMC10196665_f1",
  "image": "images/PMC10196665-f1.png",
  "conversations": [
    {
      "from": "human",
      "value": "<image>\nDescribe this scientific figure in detail."
    },
    {
      "from": "gpt",
      "value": "Figure 1. Schematic of DNAJC5 sequence alignment and dnj-14 C. elegans mutants CRISPR-Cas9..."
    }
  ]
}

Training File Descriptions

figure_caption.json - Basic figure captioning

• Single image → single detailed caption
• Best for: Training basic figure understanding

figure_detailed.json - Contextual descriptions

• Includes paper title and abstract for richer context
• Best for: Training models to understand figures in research context

figure_qa.json - Multi-turn conversations

• 3-turn Q&A: figure type → detailed description → source info
• Best for: Training conversational/interactive models

article_multiimage.json - Multi-figure reasoning

• 2-5 figures from same paper with combined analysis
• Best for: Training models to relate multiple figures

combined_training.json - Everything shuffled together

• All formats mixed for diverse training
• Recommended for most training scenarios

---

Response Length Statistics

Critical for setting max_new_tokens during training/inference:

Percentile	Characters	Est. Tokens
Median	506	~127
95th	3,383	~845
99th	6,741	~1,685
Max	22,518	~5,630

Recommended Training Configuration

# Training
model_max_length = 4096  # or 8192 for safety

# Inference - IMPORTANT: Don't set too low, may truncate responses!
generation_config = {
    "max_new_tokens": 2048,  # Covers 99th percentile
    "min_new_tokens": 100,   # Prevents cutoffs
    "do_sample": True,
    "temperature": 0.7,
}

---

Image Statistics

Metric	Value
Median dimensions	738 × 639 px
Size distribution	82% medium (500-1000px)
Tiny images (<200px)	0.5%
Format	PNG/JPG

Preprocessing recommendations:

• LLaVA: 448×448, 512x512 or 672×672
• Qwen3-VL: Dynamic resolution (native support)

---

Caption Statistics

Category	Count	Percentage
Very short (<100 chars)	14,339	8.7%
Short (100-500 chars)	43,031	26.2%
Medium (500-1500 chars)	80,614	49.1%
Long (1500-3000 chars)	24,645	15.0%
Very long (>3000 chars)	1,493	0.9%

---

Usage

Loading with Hugging Face

from datasets import load_dataset

# Load the main training file
dataset = load_dataset("SkyWhal3/STXBP1_PubMed_Central_Multimodal_Dataset", 
                       data_files="combined_training.json")

# Or load specific formats
captions = load_dataset("SkyWhal3/STXBP1_PubMed_Central_Multimodal_Dataset",
                        data_files="figure_caption.json")

Training with LLaVA

# Point to the training file
data_path = "combined_training.json"
image_folder = "images/"

# Ensure proper max_length settings
training_args = TrainingArguments(
    # ... your config
)

# Model config
model.config.max_length = 4096

Training with Qwen3-VL

# Qwen3-VL handles dynamic resolution natively
# Just ensure max_new_tokens is set properly for inference
generation_config = {
    "max_new_tokens": 2048,
}

---

About STXBP1

STXBP1 (Syntaxin-Binding Protein 1), also known as Munc18-1, is essential for synaptic vesicle fusion and neurotransmitter release. Mutations cause STXBP1 Encephalopathy, a rare neurological disorder (~1 in 30,000 births) characterized by:

• Early-onset epilepsy
• Developmental delays
• Movement disorders
• Intellectual disability

This dataset supports research into understanding and treating STXBP1-related conditions.

---

Dataset Construction

1. Source: 31,786 articles from PubMed Central related to STXBP1, synaptic function, and neurological research
2. Extraction: Custom HTML parser extracting figures, captions, abstracts, and full text
3. Matching: 99.7% of extracted figures matched to downloaded images
4. Validation: Comprehensive quality checks on text lengths and image dimensions
5. Formatting: Multiple LLaVA-compatible training formats generated

---

Citation

If you use this dataset, please cite:

@dataset{stxbp1_multimodal_2025,
  author = {SkyWhal3},
  title = {STXBP1 PubMed Central Multimodal Dataset},
  year = {2025},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/SkyWhal3/STXBP1_PubMed_Central_Multimodal_Dataset}
}

---

License

This dataset is released under CC-BY-4.0. The source articles are from PubMed Central's Open Access subset.

---

Changelog

v2.0 (December 13, 2025)

• Complete rebuild with improved extraction pipeline
• 497,360 training examples (16x increase)
• 99.7% figure-image match rate
• Full captions without truncation
• Multiple training formats (caption, detailed, Q&A, multi-image)
• Comprehensive validation and statistics

v1.0 (December 7, 2025)

• Initial release
• 31,585 articles
• Basic LLaVA/conversational/simple formats