Khalilbraham/sc-shortcutbench-private

SC-ShortcutBench: Private Test Dataset (Evaluation-Only)

⚠️ IMPORTANT: Evaluation-Only Access

This dataset contains test expression data for fair benchmark evaluation. Access is restricted to prevent overfitting.

DO NOT use this data for:

• Training your models
• Hyperparameter tuning
• Feature selection
• Any other development purposes

Dataset Overview

This is the private evaluation split of SC-ShortcutBench. It contains identical expression data to the public dataset but with restricted access to enforce proper evaluation protocols.

• Public dataset: https://huggingface.co/datasets/Khalilbraham/sc-shortcutbench-public

  • Use for: training, development, analysis
  • Access: Open to all

• Private dataset (this one): https://huggingface.co/datasets/Khalilbraham/sc-shortcutbench-private

  • Use for: final evaluation only
  • Access: Manual approval required

Files

Expression Data (Test Set)

• balanced_expression_test.h5ad (340 MB)

  • 19,250 cells (same cells as public, restricted access)
  • Balanced split with conflict-row annotations

• decorrelated_expression_test.h5ad (268 MB)

  • 14,420 cells (same cells as public, restricted access)
  • Decorrelated split with conflict-row annotations

Evaluation Protocol

Correct Usage ✅

from datasets import load_dataset
import scanpy as sc
import pandas as pd

# 1. Load PUBLIC training data
public = load_dataset("Khalilbraham/sc-shortcutbench-public")
adata_train = sc.read_h5ad("balanced_expression.h5ad")  # From public dataset

# 2. Train your model
model = train_my_model(adata_train)

# 3. Load PRIVATE test data (requires approval)
private = load_dataset("Khalilbraham/sc-shortcutbench-private", token=True)
adata_test = sc.read_h5ad("balanced_expression_test.h5ad")

# 4. Evaluate
results = evaluate_model(model, adata_test)
print(f"Truth accuracy: {results['truth_accuracy']}")
print(f"Shortcut accuracy: {results['shortcut_accuracy']}")
print(f"TSM: {results['TSM']}")

Incorrect Usage ❌

# DON'T train on private data
model.fit(adata_test)  # ❌ WRONG

# DON'T use for hyperparameter tuning
best_model = tune_hyperparameters_on_private_data()  # ❌ WRONG

# DON'T analyze to inform model development
if "pattern in private data":  # ❌ WRONG
    modify_model()

Access Request

To request access to this dataset:

1. Go to: https://huggingface.co/datasets/Khalilbraham/sc-shortcutbench-private
2. Click "Request Access"
3. Fill in:
   • Affiliation: Your institution
   • Purpose: Brief description (example: "Evaluating X model on SC-ShortcutBench benchmark")
   • Commitment: "I understand this is evaluation-only data and will not train on it"
4. Submit for approval

Approval typically takes 1-2 weeks.

Data Format

AnnData Objects

Same structure as public dataset:

• X: Gene expression matrix (log-normalized counts)
• var: Gene metadata (61,888 genes)
• obs: Cell-level metadata
  • cell_type, tissue_general, disease, dataset_id, etc.
  • tissue_ontology_term_id, cell_type_ontology_term_id
  • Donor and technical metadata

Key Columns for Evaluation

# Load test data
adata = sc.read_h5ad("balanced_expression_test.h5ad")

# Metadata for results computation
targets = {
    "cell_type": adata.obs["cell_type"],      # True labels
    "tissue_general": adata.obs["tissue_general"],
    "disease": adata.obs["disease"]
}

# For grouping (important for confidence intervals)
group = adata.obs["dataset_id"]  # Group by source study

License

CC-BY-4.0

You are free to:

• Report results on this benchmark
• Cite the dataset in your paper
• Share results publicly

You must:

• Acknowledge CELLxGENE Census
• Cite the SC-ShortcutBench paper
• Not redistribute the raw data

Citation

@article{sc_shortcutbench_2026,
  title={SC-ShortcutBench: A Conflict-Row Benchmark for Metadata Shortcut Reliance in Single-Cell Foundation Models},
  author={Your Author Names},
  journal={Proceedings of the Conference on Neural Information Processing Systems},
  year={2026},
  note={Datasets and Benchmarks Track}
}

Questions

• Dataset access: Request through HuggingFace (click Request Access)
• Technical questions: [Your contact email]
• Paper/methods: See full paper and public dataset README

Related Resources

• Public Dataset: https://huggingface.co/datasets/Khalilbraham/sc-shortcutbench-public
• Paper: [Your paper link]
• Code: [Your code repository]
• CELLxGENE Census: https://cellxgene.cziscience.com/

---

Remember: This dataset is for evaluation only. Use the public dataset for all development work.