Turn evaluator/network failures into readable errors instead of a bare 500

app.py

@@ -25,6 +25,7 @@ from __future__ import annotations
 
 import json
 import os
+import traceback
 
 import gradio as gr
 import pandas as pd
@@ -32,6 +33,13 @@ import requests
 from gradio_client import Client, handle_file
 from huggingface_hub import hf_hub_download
 
+try:
+    # gradio_client wraps a `gr.Error` raised by the backend as AppError. Guarded
+    # because a missing import here would kill the Space at startup.
+    from gradio_client.exceptions import AppError
+except ImportError:  # pragma: no cover - fall back to the class-name check below
+    AppError = None
+
 HF_TOKEN = os.environ["HF_TOKEN"]
 BACKEND = os.environ.get("BACKEND_SPACE", "nvidia/PhysicalAI-OOD-Evaluator-Private")
 RESULTS_REPO = os.environ.get("RESULTS_REPO", "nvidia/PhysicalAI-OOD-Results")
@@ -171,13 +179,33 @@ def submit(pred_file, model_name, organization, track, publish,
     if pred_file is None:
         raise gr.Error("Please upload a submission.json file.")
     progress(0.15, desc="Connecting to evaluator…")
-    client = Client(BACKEND, token=HF_TOKEN)
-    progress(0.4, desc="Running GPU evaluation (this can take a minute)…")
-    row, md = client.predict(
-        handle_file(pred_file), model_name.strip(), organization.strip(), email,
-        bool(publish), track, profile.username, str(profile.get("sub") or ""),
-        api_name="/evaluate",
-    )
+    try:
+        client = Client(BACKEND, token=HF_TOKEN)
+        progress(0.4, desc="Running GPU evaluation (this can take a minute)…")
+        row, md = client.predict(
+            handle_file(pred_file), model_name.strip(), organization.strip(), email,
+            bool(publish), track, profile.username, str(profile.get("sub") or ""),
+            api_name="/evaluate",
+        )
+    except gr.Error:
+        raise
+    except Exception as e:
+        # Anything that is not a gr.Error arrives here as a bare exception, which
+        # Gradio renders as an unparseable 500 ("Unexpected token 'I'") rather than
+        # a message -- so turn every path into something the participant can act on.
+        traceback.print_exc()
+        from_backend = (AppError is not None and isinstance(e, AppError)) or \
+            type(e).__name__ == "AppError"
+        if from_backend:
+            # The evaluator rejected the submission on purpose; its message is
+            # already written for the participant.
+            raise gr.Error(str(e)) from e
+        raise gr.Error(
+            "Could not reach the evaluator, so nothing was scored. This is on our "
+            "side, not your file — please retry in a few minutes. If it keeps "
+            "failing, report it with your HF username and the time, quoting: "
+            f"{type(e).__name__}: {e}"
+        ) from e
     progress(1.0, desc="Done")
     # Refresh both boards (a submission only changes one, but refreshing both is cheap).
     return md, load_leaderboard("track1"), load_leaderboard("track2")