update

app.py

@@ -3,7 +3,8 @@ import pandas as pd
 from huggingface_hub import snapshot_download
 
 from src.display.css_html_js import custom_css
-from src.leaderboard.student_results import get_student_leaderboard_df, get_student_status
+from src.leaderboard.student_results import get_student_leaderboard_df, get_student_status, get_latest_submission
+import time
 from src.submission.student_queue import queue_student_submission
 from src.envs import EVAL_RESULTS_PATH, RESULTS_REPO, TOKEN
 
@@ -31,10 +32,8 @@ with demo:
                     preproc_py_uploader = gr.File(label="Preprocessing (preprocess.py)", file_count="single", file_types=[".py"])
                 with gr.Row():
                     eval_btn = gr.Button("Upload and Queue Evaluation")
-                    status_btn = gr.Button("Check Status")
                 with gr.Row():
                     student_submit_result = gr.Markdown()
-                    student_status_result = gr.Markdown()
 
             def _file_path(obj):
                 if obj is None:
@@ -45,18 +44,55 @@ with demo:
                     return obj.name
                 return str(obj)
 
-            def handle_student_submit(group_id, alias, state_dict_file, model_py_file, preproc_py_file):
+            def handle_student_submit(group_id, alias, state_dict_file, model_py_file, preproc_py_file, progress=gr.Progress()):
                 try:
-                    msg = queue_student_submission(
+                    msg, ts = queue_student_submission(
                         group_id=group_id,
                         alias=alias,
                         state_dict_file=_file_path(state_dict_file),
                         model_py_file=_file_path(model_py_file),
                         preproc_py_file=_file_path(preproc_py_file),
                     )
+                    progress(0.2, desc="Queueing...")
                 except Exception as e:
-                    msg = f"Error during evaluation: {e}"
-                return msg
+                    return f"Error during submission: {e}"
+
+                # Polling loop
+                start_time = time.time()
+                while True:
+                    time.sleep(5)
+                    if time.time() - start_time > 600: # 10 min timeout
+                        return f"Timed out waiting for results. Please check back later. Submission ID: {ts}"
+                    
+                    try:
+                        snapshot_download(
+                            repo_id=RESULTS_REPO,
+                            local_dir=EVAL_RESULTS_PATH,
+                            repo_type="dataset",
+                            tqdm_class=None,
+                            etag_timeout=30,
+                            token=TOKEN,
+                        )
+                    except Exception:
+                        pass # Ignore network errors during polling
+                    
+                    submission = get_latest_submission(group_id)
+                    if not submission:
+                        continue
+                        
+                    # Check if this is the submission we just made
+                    if submission.get("timestamp") != ts:
+                        continue
+                        
+                    status = submission.get("status", "UNKNOWN")
+                    if status == "PENDING":
+                        progress(0.5, desc="Evaluating...")
+                    elif status == "SUCCESS":
+                        progress(1.0, desc="Done!")
+                        return "Evaluation Successful! Check the Leaderboard tab."
+                    elif status == "FAILED":
+                        progress(1.0, desc="Failed")
+                        return f"Submission Failed: {submission.get('error', 'Unknown error')}"
 
             eval_btn.click(
                 handle_student_submit,
@@ -64,11 +100,6 @@ with demo:
                 [student_submit_result],
             )
 
-            status_btn.click(
-                get_student_status,
-                [group_id_tb],
-                [student_status_result],
-            )
 
         with gr.TabItem("📊 Leaderboard", elem_id="student-leaderboard-tab"):
             # Ensure latest results are cached locally at startup

src/leaderboard/student_results.py

@@ -134,3 +134,29 @@ def get_student_status(group_id: str) -> str:
         return "" # User requested to say nothing if success
     else:
         return f"Status: {status}"
+
+
+def get_latest_submission(group_id: str) -> Optional[Dict]:
+    attempts_dir = os.path.join(EVAL_RESULTS_PATH, PROJ_DIR, "student_attempts")
+    if not os.path.isdir(attempts_dir):
+        return None
+
+    rows = []
+    for fname in os.listdir(attempts_dir):
+        if not fname.endswith(".json"):
+            continue
+        try:
+            with open(os.path.join(attempts_dir, fname), "r") as f:
+                data = json.load(f)
+                # Ensure group_id comparison is robust (string vs int)
+                if str(data.get("group_id", "")).strip() == str(group_id).strip():
+                    rows.append(data)
+        except Exception:
+            continue
+
+    if not rows:
+        return None
+
+    # Sort by timestamp desc
+    rows.sort(key=lambda x: x.get("timestamp", ""), reverse=True)
+    return rows[0]

src/submission/student_queue.py

@@ -22,7 +22,7 @@ def queue_student_submission(
     state_dict_file: Optional[str],
     model_py_file: str,
     preproc_py_file: str,
-) -> str:
+) -> tuple[str, str]:
     """
     Uploads submitted files to the private queue dataset for offline evaluation.
     Layout in repo:
@@ -97,6 +97,6 @@ def queue_student_submission(
     except Exception:
         pass
 
-    return f"Submission queued for Group '{group_id}' at {ts}. Your model will be evaluated shortly."
+    return f"Submission queued for Group '{group_id}' at {ts}. Your model will be evaluated shortly.", ts