Nicer intro: 3-step submit flow + data download link

app.py

@@ -41,14 +41,15 @@ DETAIL_COLUMNS = ["dataset_id", "status", "skill"]
 
 def _read_results() -> pd.DataFrame:
     from huggingface_hub import hf_hub_download
+    from huggingface_hub.utils import EntryNotFoundError, RepositoryNotFoundError
 
     try:
         path = hf_hub_download(
             RESULTS_REPO, RESULTS_FILE, repo_type="dataset", token=TOKEN
         )
-        return pd.read_csv(path)
-    except Exception:
+    except (RepositoryNotFoundError, EntryNotFoundError):
         return pd.DataFrame(columns=RESULT_COLUMNS)
+    return pd.read_csv(path)
 
 
 def _append_results(rows: list[dict]) -> None:
@@ -125,8 +126,8 @@ def leaderboard() -> pd.DataFrame:
     joined from the task registry at read time; if that fetch fails the board
     renders empty rather than showing a wrong ranking.
     """
-    df = _read_results()
     try:
+        df = _read_results()
         datasets = manifest_ids(fetch_manifest(TOKEN))
         registry = scoreable_tasks(fetch_tasks_registry(TOKEN), datasets)
         by_id = {_norm_id(task["task_id"]): task for task in registry}
@@ -258,14 +259,21 @@ def evaluate(submission_path: str, model_name: str):
 def build_demo() -> gr.Blocks:
     with gr.Blocks(title="Lodestar Benchmark") as demo:
         gr.Markdown(
-            "# 🧬 Lodestar\n"
-            "Blind benchmark for transcriptomic foundation models. Embed **every** "
-            "dataset and upload **one** file; a fixed linear probe scores each dataset "
-            "and rolls up into a per-facet skill leaderboard. Only full-coverage "
-            "submissions are ranked.\n\n"
-            "Submission = CSV / TSV / Parquet with a `dataset_id` column, a `sample_id` "
-            "column, and one column per embedding dim; or NPZ with `dataset_ids`, "
-            "`sample_ids`, `embeddings` arrays."
+            "# 🧬 Lodestar\n\n"
+            "A **blind benchmark for transcriptomic foundation models** — grade "
+            "your model's patient-level embeddings against real clinical signal, "
+            "without ever seeing the labels.\n\n"
+            "**Submit in 3 steps:**\n"
+            "1. **Get the data** → download the opaque datasets from "
+            "[ScientaLab/lodestar](https://huggingface.co/datasets/ScientaLab/lodestar)"
+            " (start with its `datasets.yaml`).\n"
+            "2. **Embed every dataset** → build **one** file: `dataset_id`, "
+            "`sample_id`, then one column per embedding dim (`e0`, `e1`, …). "
+            "CSV / TSV / Parquet, or NPZ.\n"
+            "3. **Upload below**, name your model, and hit **Evaluate**.\n\n"
+            "A fixed linear probe scores each hidden task (AUROC or Pearson), "
+            "rescaled to a 0–1 skill and rolled up per specialty. **Only "
+            "submissions covering every dataset are ranked.**"
         )
         model_tb = gr.Textbox(label="Model name", placeholder="e.g. eva-rna-v1")
         file_in = gr.File(