SetFit few-shot classifier for identifying eye imaging datasets from scientific metadata.
Part of the EyeACT project by the FAIR Data Innovations Hub.
pip install envision-classifierfrom envision_classifier import EyeImagingClassifier
# Downloads model from HuggingFace on first use
clf = EyeImagingClassifier()
# Classify a single record
result = clf.classify("Retinal OCT dataset for diabetic retinopathy")
print(result)
# {'label': 'EYE_IMAGING', 'confidence': 0.98,
# 'probabilities': {'EYE_IMAGING': 0.98, 'NEGATIVE': 0.02}}
# Classify a batch
results = clf.classify_batch([
"Retinal fundus photography dataset for glaucoma screening",
"COVID-19 genome sequencing data",
{"title": "OCT images", "description": "Macular degeneration scans"},
])
# Use a local model instead of downloading
clf = EyeImagingClassifier(model_path="./my_model")After installing, the envision-classifier command is available:
# Classify a text string
envision-classifier classify --text "Retinal OCT dataset for diabetic retinopathy"
# Classify from a JSON file
envision-classifier classify records.json
# Pipe JSON via stdin
echo '{"title": "Fundus images", "description": "DR screening"}' | envision-classifier classify
# Train a new model from built-in training data
envision-classifier train --output ./my_model
# Show model info and training data counts
envision-classifier info| Label | Description |
|---|---|
| EYE_IMAGING | Actual eye imaging datasets (fundus, OCT, OCTA, cornea) |
| NEGATIVE | Everything else (software, non-imaging eye data, unrelated domains) |
- Base model:
sentence-transformers/all-mpnet-base-v2(768-dim) - Training data: 891 curated examples (262 EYE_IMAGING, 629 NEGATIVE) from Zenodo, Figshare, Dryad, Kaggle, and NEI
- Test accuracy: 0.961, EYE_IMAGING F1: 0.936
- Spot-check: 30/33 (90.9%)
- Model weights: fairdataihub/envision-eye-imaging-classifier
Applied across 6 repositories via envision-discovery:
| Source | EYE_IMAGING | NEGATIVE | Total |
|---|---|---|---|
| Zenodo | 60 | 455 | 515 |
| DataCite | 752 | 1,084 | 1,836 |
| Figshare | 1,049 | 951 | 2,000 |
| Kaggle | 248 | 484 | 732 |
| Dryad | 32 | 57 | 89 |
| NEI | 686 | 976 | 1,662 |
Classification is based on metadata only (titles, descriptions, keywords, and file types inspected inside archives via HTTP Range requests) -- no dataset files are downloaded.
- envision-discovery -- Full pipeline (scraping + classification + export)
- Model on HuggingFace
MIT