fix: show 5-segment prompt (level_3) with segment headers

app.py

@@ -67,12 +67,23 @@ TASK_BY_ID: dict[str, dict[str, Any]] = {t["task_id"]: t for t in TASKS}
 
 
 def _extract_prompt(task: dict[str, Any]) -> str:
+    """Return the 5-segment prompt (level_3), joined with segment headers.
+
+    All MBench-V models were actually fed the 5-segment (level_3) prompts at
+    inference time, so this is what annotators should see. If level_3 is
+    missing for some reason we fall back to the finest available level.
+    """
     gp = task.get("generation_prompts") or {}
     prompts = gp.get("prompts") or {}
-    for level in ("level_1", "level_2", "level_3"):
+    # Prefer the canonical 5-segment split; otherwise fall back gracefully.
+    for level in ("level_3", "level_4", "level_2", "level_1"):
         val = prompts.get(level)
         if isinstance(val, list) and val:
-            return val[0]
+            n = len(val)
+            parts = []
+            for i, seg in enumerate(val, 1):
+                parts.append(f"— Segment {i}/{n} —\n{seg}")
+            return "\n\n".join(parts)
         if isinstance(val, str) and val:
             return val
     return "(no prompt found)"
@@ -367,9 +378,9 @@ with gr.Blocks(title="MBench-V Annotation", theme=gr.themes.Soft()) as demo:
         with gr.Column(scale=2):
             meta_md = gr.Markdown()
             prompt_tb = gr.Textbox(
-                label="Generation prompt",
-                lines=10,
-                max_lines=20,
+                label="Generation prompt (5 segments)",
+                lines=20,
+                max_lines=40,
                 interactive=False,
             )
         with gr.Column(scale=1):