PubGuard — Multi-Head Scientific Publication Gatekeeper

Model Description

PubGuard is a lightweight, CPU-optimized document classifier that screens PDF text to determine whether it represents a genuine scientific publication. It runs as Step 0 in the PubVerse + 42DeepThought pipeline, rejecting non-publications (posters, abstracts, review articles, flyers, invoices) before expensive downstream processing (VLM feature extraction, graph construction, GNN scoring).

Three classification heads provide a multi-dimensional screening verdict:

Document type — Is this a paper, review, poster, abstract, or junk?
AI detection — Was this written by a human or generated by an LLM?
Toxicity — Does this contain toxic or offensive content?

Developed by Jamey O'Neill at the California Medical Innovations Institute (CalMI²).

Architecture

Three linear classification heads on frozen model2vec (potion-base-32M) embeddings:

┌─────────────┐
│  PDF text    │
└──────┬──────┘
       │
┌──────▼──────┐     ┌───────────────────┐
│  clean_text │────►│  model2vec encode  │──► emb ∈ R^512
└─────────────┘     └───────────────────┘
                            │
          ┌─────────────────┼─────────────────┐
          ▼                 ▼                  ▼
┌─────────────────┐ ┌──────────────┐ ┌──────────────┐
│ doc_type head    │ │ ai_detect    │ │ toxicity     │
│ [emb + 14 feats] │ │ head         │ │ head         │
│ → softmax(5)     │ │ → softmax(2) │ │ → softmax(2) │
└─────────────────┘ └──────────────┘ └──────────────┘

Each head is a single linear layer stored as a numpy .npz file (8–12 KB). Inference is pure numpy — no torch needed at prediction time.

The doc_type head additionally receives 14 structural features (section headings present, citation density, sentence length, etc.) concatenated with the embedding — these act as strong Bayesian priors.

Performance

Head	Classes	Accuracy	F1
doc_type	5	94.4%	0.944
ai_detect	2	84.2%	0.842
toxicity	2	83.9%	0.839

doc_type Breakdown

Class	Precision	Recall	F1
scientific_paper	0.891	0.932	0.911
literature_review	0.914	0.884	0.899
poster	0.938	0.917	0.928
abstract_only	0.985	0.991	0.988
junk	0.992	0.996	0.994

Throughput

302 docs/sec single-document, 568 docs/sec batched (CPU only)
3.3ms per PDF screening — negligible pipeline overhead
No GPU required

Gate Logic

Only scientific_paper passes the gate. Everything else — literature reviews, posters, standalone abstracts, junk — is blocked. The PubVerse pipeline processes original research publications only.

scientific_paper   →  ✅ PASS
literature_review  →  ❌ BLOCKED  (narrative/scoping reviews)
poster             →  ❌ BLOCKED  (classified, but not a publication)
abstract_only      →  ❌ BLOCKED
junk               →  ❌ BLOCKED

Note: Meta-analyses and systematic reviews are classified as scientific_paper (they are primary research). Only narrative and scoping reviews are classified as literature_review.

AI detection and toxicity are informational by default — reported but not blocking.

Usage

Python API

from pubguard import PubGuard

guard = PubGuard()
guard.initialize()

verdict = guard.screen("Introduction: We present a novel deep learning approach...")
print(verdict)
# {
#   'doc_type': {'label': 'scientific_paper', 'score': 0.994},
#   'ai_generated': {'label': 'human', 'score': 0.875},
#   'toxicity': {'label': 'clean', 'score': 0.999},
#   'pass': True
# }

Pipeline Integration (bash)

# Step 0 in run_pubverse_pipeline.sh:
PDF_TEXT=$(python3 -c "import fitz; d=fitz.open('$pdf'); print(' '.join(p.get_text() for p in d)[:8000])")
PUBGUARD_CODE=$(echo "$PDF_TEXT" | python3 pub_check/scripts/pubguard_gate.py 2>/dev/null)
# exit 0 = pass, exit 1 = reject

Installation

pip install git+https://github.com/jimnoneill/pubguard.git

With training dependencies:

pip install "pubguard[train] @ git+https://github.com/jimnoneill/pubguard.git"

Training Data

Trained on real datasets — zero synthetic data:

Head	Sources	Samples
doc_type	PDF corpus (microbiome/metagenomics), armanc/scientific_papers, OpenAlex OA review PDFs + abstracts, poster-sentry-training-data, gfissore/arxiv-abstracts-2021, ag_news	75K (15K per class)
ai_detect	liamdugan/raid (abstracts), NicolaiSivesind/ChatGPT-Research-Abstracts	30K
toxicity	google/civil_comments, skg/toxigen-data	30K

The poster class uses real scientific poster text from the posters.science corpus (28K+ verified posters from Zenodo & Figshare), extracted by PosterSentry.

The literature_review class uses a mix of open-access review article PDFs downloaded from OpenAlex and review abstracts as fallback. See the training data repo for full details.

Training

python scripts/train_pubguard.py --pdf-corpus /path/to/pdfs --n-per-class 15000

Training completes in ~1 minute on CPU. No GPU needed.

Model Specifications

Attribute	Value
Embedding backbone	minishlab/potion-base-32M (model2vec StaticModel)
Embedding dimension	512
Structural features	14 (doc_type head only)
Classifier	LogisticRegression (sklearn) per head
Head file sizes	5–12 KB each (.npz)
Total model size	~125 MB (embedding) + 25 KB (heads)
Precision	float32
GPU required	No (CPU-only)
License	MIT

Citation

@software{pubguard_2026,
  title = {PubGuard: Multi-Head Scientific Publication Gatekeeper},
  author = {O'Neill, James},
  year = {2026},
  url = {https://huggingface.co/jimnoneill/pubguard-classifier},
  note = {Part of the PubVerse + 42DeepThought pipeline}
}

License

This model is released under the MIT License.

Acknowledgments

California Medical Innovations Institute (CalMI²)
MinishLab for the model2vec embedding backbone
FAIR Data Innovations Hub for the PosterSentry training data
HuggingFace for model hosting infrastructure

Downloads last month: 37