Deploy Space: update app.py

app.py

@@ -22,6 +22,10 @@ from src.utils import (
     load_baseline_rank_history,
     load_baseline_rank_history_weekly,
     load_forecast_snapshots,
+    load_gift_aggregate_metadata_md,
+    load_gift_aggregate_table,
+    load_live_aggregate_metadata_md,
+    load_live_aggregate_table,
     make_domain_pie_chart,
     make_forecast_plot,
     make_rank_trend_plot,
@@ -31,6 +35,7 @@ from src.utils import (
 
 RESULTS_PATH = os.getenv("TSFM_RESULTS_PATH", "space/results" if Path("space/results").exists() else "results")
 REFRESH_SECONDS = int(os.getenv("TSFM_REFRESH_SECONDS", "300"))
+DOMAIN_OUTPUT_OFFSET = 13
 
 
 def _dataset_tables(grouped: dict, datasets: list[str]) -> tuple[list, list]:
@@ -81,16 +86,23 @@ def refresh_leaderboard() -> tuple:
         daily_df,                                           # [3] daily table
         load_baseline_rank_history_weekly(RESULTS_PATH),    # [4] weekly table
         make_rank_trend_plot(daily_df),                     # [5] rank trend plot
+        load_gift_aggregate_metadata_md(RESULTS_PATH),      # [6]
+        load_gift_aggregate_table(RESULTS_PATH, "prediction_length"),  # [7]
+        load_gift_aggregate_table(RESULTS_PATH, "domain"),             # [8]
+        load_gift_aggregate_table(RESULTS_PATH, "frequency"),          # [9]
+        load_live_aggregate_metadata_md(RESULTS_PATH),                 # [10]
+        load_live_aggregate_table(RESULTS_PATH, "overall"),           # [11]
+        load_live_aggregate_table(RESULTS_PATH, "rank"),              # [12]
     ]
 
-    # [6 .. 6+D-1] per-domain status lines
+    # [10 .. 10+D-1] per-domain status lines
     for domain, domain_datasets in DOMAIN_GROUPS.items():
         outputs.append(build_domain_status_md(domain, domain_datasets, RESULTS_PATH))
 
     value_tables, rank_tables = _dataset_tables(grouped, datasets)
-    # [6+D .. 6+D+N-1] value tables
+    # [10+D .. 10+D+N-1] value tables
     outputs.extend(value_tables)
-    # [6+D+N .. 6+D+2N-1] rank tables
+    # [10+D+N .. 10+D+2N-1] rank tables
     outputs.extend(rank_tables)
     return tuple(outputs)
 
@@ -117,6 +129,30 @@ with demo:
     with gr.Tabs(elem_classes="tab-buttons"):
         # ── Overall Tab ──────────────────────────────────────────────────────
         with gr.TabItem("Overall"):
+            live_aggregate_metadata_md = gr.Markdown(
+                INITIAL_OUTPUTS[10],
+                elem_classes="markdown-text",
+            )
+            gr.Markdown(
+                "**Cumulative live evaluation.** MSE, RMSE, MAPE, CRPS, and Stability are lower-is-better; RTG is higher-is-better; more negative Improvement indicates a stronger decreasing-error trend.",
+                elem_classes="markdown-text",
+            )
+            live_overall_df = gr.Dataframe(
+                value=INITIAL_OUTPUTS[11],
+                interactive=False,
+                wrap=True,
+                label="Up-to-now overall results",
+            )
+            live_rank_df = gr.Dataframe(
+                value=INITIAL_OUTPUTS[12],
+                interactive=False,
+                wrap=True,
+                label="Average Rank, Win Rate, and Elo",
+            )
+            gr.Markdown(
+                "**Latest evaluation snapshot**",
+                elem_classes="markdown-text",
+            )
             gr.Markdown(
                 "**Evaluated timestamps by domain** · hover for dataset names",
                 elem_classes="markdown-text",
@@ -171,6 +207,39 @@ with demo:
                 label="Rank trend over time",
             )
 
+        # ── GIFT-style grouped result tables ────────────────────────────────
+        with gr.TabItem("GIFT-style Aggregates"):
+            aggregate_metadata_md = gr.Markdown(
+                INITIAL_OUTPUTS[6],
+                elem_classes="markdown-text",
+            )
+            gr.Markdown(
+                "MSE and CRPS are normalized per dataset configuration against Seasonal-Naive (1.0 = baseline). Rank is the mean per-configuration CRPS rank. Lower is better for all three metrics.",
+                elem_classes="markdown-text",
+            )
+            with gr.Tabs():
+                with gr.TabItem("Prediction Length"):
+                    prediction_length_aggregate_df = gr.Dataframe(
+                        value=INITIAL_OUTPUTS[7],
+                        interactive=False,
+                        wrap=True,
+                        label="Results on TSFM_Bench aggregated by Prediction Length",
+                    )
+                with gr.TabItem("Domain"):
+                    domain_aggregate_df = gr.Dataframe(
+                        value=INITIAL_OUTPUTS[8],
+                        interactive=False,
+                        wrap=True,
+                        label="Results on TSFM_Bench aggregated by Domain",
+                    )
+                with gr.TabItem("Frequency"):
+                    frequency_aggregate_df = gr.Dataframe(
+                        value=INITIAL_OUTPUTS[9],
+                        interactive=False,
+                        wrap=True,
+                        label="Results on TSFM_Bench aggregated by Frequency",
+                    )
+
         # ── Domain Tabs ───────────────────────────────────────────────────────
         domain_idx = 0
         for domain, domain_datasets in DOMAIN_GROUPS.items():
@@ -178,7 +247,7 @@ with demo:
                 # Per-domain status line (data source, last eval, refresh intervals)
                 domain_status_mds.append(
                     gr.Markdown(
-                        INITIAL_OUTPUTS[6 + domain_idx],
+                        INITIAL_OUTPUTS[DOMAIN_OUTPUT_OFFSET + domain_idx],
                         elem_classes="markdown-text",
                     )
                 )
@@ -191,7 +260,7 @@ with demo:
                     )
                     dataset_value_dfs.append(
                         gr.Dataframe(
-                            value=INITIAL_OUTPUTS[6 + N_DOMAINS + dataset_idx],
+                            value=INITIAL_OUTPUTS[DOMAIN_OUTPUT_OFFSET + N_DOMAINS + dataset_idx],
                             interactive=False,
                             wrap=True,
                             show_label=False,
@@ -203,7 +272,7 @@ with demo:
                     )
                     dataset_rank_dfs.append(
                         gr.Dataframe(
-                            value=INITIAL_OUTPUTS[6 + N_DOMAINS + len(DATASETS) + dataset_idx],
+                            value=INITIAL_OUTPUTS[DOMAIN_OUTPUT_OFFSET + N_DOMAINS + len(DATASETS) + dataset_idx],
                             interactive=False,
                             wrap=True,
                             show_label=False,
@@ -280,6 +349,13 @@ with demo:
         baseline_history_df,
         baseline_weekly_df,
         rank_trend_plot,
+        aggregate_metadata_md,
+        prediction_length_aggregate_df,
+        domain_aggregate_df,
+        frequency_aggregate_df,
+        live_aggregate_metadata_md,
+        live_overall_df,
+        live_rank_df,
         *domain_status_mds,
         *dataset_value_dfs,
         *dataset_rank_dfs,