---
pretty_name: "Lodestar — benchmark inputs"
license: other
---

# 🧬 Lodestar — benchmark inputs

**Lodestar** is a blind benchmark for transcriptomic foundation models: it measures
how well a model's patient-level embeddings capture real clinical signal.

This repo holds the **inputs you embed**. It's deliberately *blind* — datasets are
named `d001`, `d002`, … with no disease, tissue, or target revealed. You grade the
embedding, not task-specific tuning.

## 📦 What's inside
- **`datasets.yaml`** — the manifest: each dataset's `id`, file `path`, and shape
  (`n_samples`, `n_genes`, `gene_id_type`).
- **`d001/expression.h5ad`, `d002/…`** — raw counts as `AnnData`: rows = samples
  (`sample_id`), columns = NCBI gene ids (with `gene_symbols`).

## 🚀 Run the benchmark
1. **Download** the data:
   ```bash
   hf download ScientaLab/lodestar --repo-type dataset --local-dir lodestar
   ```
2. **Embed every dataset** with your model — one vector per sample (any dimension;
   it may differ per dataset).
3. **Assemble one submission file** covering all datasets — `dataset_id`,
   `sample_id`, then one column per embedding dim (`e0`, `e1`, …). Format:
   **CSV / TSV / Parquet**, or **NPZ** with `dataset_ids` / `sample_ids` / `embeddings`.
4. **Submit & see your rank** →
   **[🏆 Lodestar Space](https://huggingface.co/spaces/ScientaLab/lodestar-eval)**

## 📊 How it's scored
A fixed linear probe is trained on frozen cross-validation folds over your embeddings
and scored — **AUROC** (classification) or **Pearson r** (regression). Scores become a
0–1 **skill** and roll up per medical specialty. **Only submissions that cover every
dataset are ranked.** Labels stay private — grading happens server-side.

➡️ **Ready?** → **https://huggingface.co/spaces/ScientaLab/lodestar-eval**

*Each dataset is redistributed under its source's original open license.*