description

main.py

@@ -87,6 +87,40 @@ with gr.Blocks(title="Impermanent Leaderboard") as app:
         "on data they could not have seen during training."
     )
 
+    cutoff_dates = sorted(df["cutoff"].unique())
+    n_dates = len(cutoff_dates)
+    date_min, date_max = cutoff_dates[0], cutoff_dates[-1]
+    statistical_models = ["zero_model", "seasonal_naive", "auto_arima", "auto_ets", "auto_lgbm"]
+    foundation_models = ["chronos", "moirai", "timesfm"]
+    all_model_names = statistical_models + foundation_models
+
+    gr.Markdown(f"""\
+## Datasets
+
+GitHub repositories are selected across several **buckets based on their number of stars**,
+yielding a mix of both intermittent (low-activity) and high-volume time series.
+For each bucket, an automated pipeline fetches four signals:
+
+- **Open issues** — number of issues opened
+- **Opened PRs** — number of pull requests opened
+- **Pushes** — number of push events
+- **Stars** — number of new stars
+
+Each signal is collected at both **daily** and **weekly** granularity.
+
+## Models
+
+The benchmark evaluates two families of forecasting methods:
+
+- **Statistical / ML models:** {", ".join(f"`{m}`" for m in statistical_models)}
+- **Foundation models:** {", ".join(f"`{m}`" for m in foundation_models)}
+
+## Evaluation dates
+
+Forecast methods are evaluated **every week** using rolling forecast evaluations.
+Currently **{n_dates} evaluations** are available, from **{date_min}** to **{date_max}**.
+""")
+
     with gr.Tab("Leaderboard 🏆"):
         lb = compute_leaderboard(df)
         gr.Dataframe(