saurabinfocusp/human-CTCL-tissue

Model Card: cell_type

Summary

• Task: cell_type_annotation
• Model version: 1.0.0
• scaLR version: 2.0.0.dev0
• Species: not specified
• Random seed: 42

Training data

• Labels column: cell_type
• Grouping column (leakage-safe split): donor_id
• Split ratio (train/val/test): [0.7, 0.1, 0.2]
• Feature count: 15675
• Classes (10): dermis lymphatic vessel endothelial cell, endothelial cell of vascular tree, fibroblast, keratinocyte, mature B cell, mature alpha-beta T cell, melanocyte of skin, pericyte, plasma cell, professional antigen presenting cell

Performance (held-out test split)

• macro-F1: 0.9629
• weighted-F1: 0.9820
• balanced accuracy: 0.9701
• test cells: 324,132

Confidence & abstention

• Calibration: temperature scaling (T=2.547)
• Open-set thresholds: {'min_confidence': 0.9546103924512863, 'max_entropy': 0.09411545395851133, 'min_margin': 0.9238014698028565}

Known limitations

• Evaluated only on the training-distribution data described above; performance on out-of-distribution donors, tissues, or sequencing technologies is not validated.
• Open-set (is_unknown) thresholds are fit on a single validation split and are a heuristic screening signal, not a guarantee of correct abstention.
• Gene/feature coverage should be checked (scalr.validate) on every new query dataset.

License

• Source code license: see repository LICENSE.
• This model artifact's weights/data license: not specified by this template; set explicitly by the model's author/distributor before sharing.