Upload app.py with huggingface_hub

app.py

@@ -109,6 +109,10 @@ def vqa(image, question, max_new_tokens=256):
         model, processor, image_paths=[path], question=question.strip(),
         device=DEVICE, dtype=DTYPE, max_new_tokens=int(max_new_tokens),
     )
+    # The model wraps answers in <answer>...</answer>; strip the tags for display.
+    text = text.strip()
+    for tag in ("<answer>", "</answer>"):
+        text = text.replace(tag, "")
     return text.strip()
 
 
@@ -197,7 +201,15 @@ def embodied_plan(image, task, num_frames=3, num_steps=50, height=144, width=256
         pil.save(frame_f.name)
         prior_steps.append((text, frame_f.name))
         gallery.append((pil, f"Step {k + 1}"))
-        plan_md_lines.append(f"**Step {k + 1}:** {text.strip() or '(imagined goal frame only)'}")
+        # Tidy the emitted text: drop any wrapper tags and a leading "Step N:"
+        # the model may already produce (we add our own numbered heading).
+        disp = text.strip()
+        for tag in ("<answer>", "</answer>"):
+            disp = disp.replace(tag, "")
+        disp = disp.strip()
+        if disp.lower().startswith(f"step {k + 1}:"):
+            disp = disp[len(f"step {k + 1}:"):].strip()
+        plan_md_lines.append(f"**Step {k + 1}:** {disp or '(imagined goal frame only)'}")
         yield gallery, "\n\n".join(plan_md_lines)