fix: simplify UI to single page; no visibility toggling; plain-value returns

app.py

@@ -1,19 +1,8 @@
 """MBench-V annotation UI (Gradio Space).
 
-Reads videos streaming from the `studyOverflow/TempMemoryData` dataset repo,
-writes annotations back to the same repo under `annotations/`, batched via
+Streams videos from `studyOverflow/TempMemoryData` (no local copy); writes
+annotations back to the same dataset repo under `annotations/`, batched via
 `CommitScheduler`.
-
-Design notes
-------------
-- Videos are NOT copied into this Space. We build CDN URLs with
-  `hf_hub_url(..., repo_type="dataset")` and let the browser stream them.
-- Submissions are appended to a per-process JSONL file under `annotations/`;
-  `CommitScheduler` pushes the directory to the dataset repo every 5 min.
-- Allocation is intentionally simple in this template: at start-up we build
-  a single shuffled pool of `(model, task_id)` pairs, and each user session
-  maintains its own index into that pool. Multi-annotator deduplication is
-  out of scope for the first iteration.
 """
 
 from __future__ import annotations
@@ -36,8 +25,8 @@ from huggingface_hub import CommitScheduler, hf_hub_download, hf_hub_url
 DATASET_REPO = "studyOverflow/TempMemoryData"
 MERGED_JSON_PATH = "MBench-V/merged.json"
 
-# 6 models that are already fully reorganized on HF (584 videos each).
-# `skyreels` and `longcat` are excluded until their 0422 runs finish.
+# 6 fully-reorganized models (584 videos each). `skyreels` and `longcat`
+# are temporarily excluded until their 0422 runs finish.
 MODELS: list[str] = [
     "causal_forcing",
     "self_forcing",
@@ -47,22 +36,17 @@ MODELS: list[str] = [
     "memflow",
 ]
 
-HF_TOKEN = os.environ.get("HF_TOKEN")  # must be set in Space secrets for writes
+HF_TOKEN = os.environ.get("HF_TOKEN")
 
-# Local staging directory that CommitScheduler will sync to the dataset repo.
 ANN_DIR = Path("annotations_local")
 ANN_DIR.mkdir(exist_ok=True)
-
-# Each Space process writes to its own JSONL so concurrent replicas don't
-# clobber each other's writes. `CommitScheduler` pushes the whole directory.
 PROCESS_ID = uuid.uuid4().hex[:8]
 ANN_FILE = ANN_DIR / f"ann_{PROCESS_ID}.jsonl"
-
 COMMIT_INTERVAL_MIN = 5
 
 
 # ---------------------------------------------------------------------------
-# Load merged.json (584 task records) once at startup
+# Load merged.json once at startup
 # ---------------------------------------------------------------------------
 
 def _load_merged() -> list[dict[str, Any]]:
@@ -81,7 +65,6 @@ 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 first non-empty prompt string found in the task record."""
     gp = task.get("generation_prompts") or {}
     prompts = gp.get("prompts") or {}
     for level in ("level_1", "level_2", "level_3"):
@@ -93,19 +76,7 @@ def _extract_prompt(task: dict[str, Any]) -> str:
     return "(no prompt found)"
 
 
-# ---------------------------------------------------------------------------
-# Build the (model, task_id) pool
-# ---------------------------------------------------------------------------
-
-def _build_pool() -> list[tuple[str, str]]:
-    pool: list[tuple[str, str]] = []
-    for m in MODELS:
-        for t in TASKS:
-            pool.append((m, t["task_id"]))
-    return pool
-
-
-POOL: list[tuple[str, str]] = _build_pool()
+POOL: list[tuple[str, str]] = [(m, t["task_id"]) for m in MODELS for t in TASKS]
 print(f"[mbench-ann] loaded {len(TASKS)} tasks × {len(MODELS)} models = {len(POOL)} items")
 
 
@@ -118,7 +89,7 @@ def _video_url(model: str, task_id: str) -> str:
 
 
 # ---------------------------------------------------------------------------
-# CommitScheduler — pushes annotations_local/ to DATASET_REPO every 5 min
+# CommitScheduler
 # ---------------------------------------------------------------------------
 
 scheduler: CommitScheduler | None = None
@@ -135,7 +106,7 @@ if HF_TOKEN:
     )
     print(f"[mbench-ann] CommitScheduler started (every {COMMIT_INTERVAL_MIN} min)")
 else:
-    print("[mbench-ann] WARNING: HF_TOKEN not set — annotations will stay local only")
+    print("[mbench-ann] WARNING: HF_TOKEN not set — annotations stay local only")
 
 
 def _append_annotation(record: dict[str, Any]) -> None:
@@ -167,7 +138,6 @@ def _format_meta(model: str, task: dict[str, Any], idx: int, total: int) -> str:
 
 
 def _load_item(pool_order: list[int], idx: int) -> tuple[str, str, str]:
-    """Return (video_url, meta_markdown, prompt_text) for position `idx`."""
     if idx < 0 or idx >= len(pool_order):
         return "", "**All done!** No more items.", ""
     model, task_id = POOL[pool_order[idx]]
@@ -180,53 +150,25 @@ def _load_item(pool_order: list[int], idx: int) -> tuple[str, str, str]:
 
 
 # ---------------------------------------------------------------------------
-# Gradio callbacks
+# Gradio callbacks — all return plain Python values (no gr.update mix)
 # ---------------------------------------------------------------------------
 
-def start_session(annotator: str, state: dict | None):
+def start_session(annotator: str, state: dict):
     annotator = (annotator or "").strip()
     if not annotator:
-        return (
-            state,
-            gr.update(visible=True),  # login panel stays
-            gr.update(visible=False),  # annotation panel hidden
-            "",
-            "",
-            "",
-            gr.update(value="Please enter a name first."),
-        )
-    # Build this user's shuffled order
+        return state, "", "⚠️ Please enter a name first.", "", "⚠️ Please enter a name first."
     order = list(range(len(POOL)))
     rng = random.Random(f"{annotator}-{int(time.time())}")
     rng.shuffle(order)
     state = {"annotator": annotator, "order": order, "idx": 0}
     video, meta, prompt = _load_item(order, 0)
-    return (
-        state,
-        gr.update(visible=False),
-        gr.update(visible=True),
-        video,
-        meta,
-        prompt,
-        gr.update(value=f"Logged in as `{annotator}`"),
-    )
+    status = f"✅ Logged in as `{annotator}` — {len(order)} items to annotate."
+    return state, video, meta, prompt, status
 
 
-def _advance(state: dict, record_submitted: bool):
-    state["idx"] += 1
-    video, meta, prompt = _load_item(state["order"], state["idx"])
-    status = (
-        f"Submitted ({state['idx']} done). Next →"
-        if record_submitted
-        else f"Skipped. Next →"
-    )
-    # Reset score + note controls
-    return state, video, meta, prompt, 3, "", status
-
-
-def submit_and_next(state: dict, score: int, note: str):
-    if state is None or state.get("idx") is None:
-        return state, "", "", "", 3, "", "Not logged in."
+def submit_and_next(state: dict, score: float, note: str):
+    if not state or "order" not in state:
+        return state, "", "⚠️ Please log in first.", "", 3, "", "⚠️ Not logged in."
     order = state["order"]
     idx = state["idx"]
     if idx >= len(order):
@@ -243,82 +185,83 @@ def submit_and_next(state: dict, score: int, note: str):
         "note": (note or "").strip(),
     }
     _append_annotation(record)
-    return _advance(state, record_submitted=True)
+    state["idx"] = idx + 1
+    video, meta, prompt = _load_item(state["order"], state["idx"])
+    return state, video, meta, prompt, 3, "", f"✅ Submitted ({state['idx']}). Next →"
 
 
 def skip_and_next(state: dict):
-    if state is None or state.get("idx") is None:
-        return state, "", "", "", 3, "", "Not logged in."
-    return _advance(state, record_submitted=False)
+    if not state or "order" not in state:
+        return state, "", "⚠️ Please log in first.", "", 3, "", "⚠️ Not logged in."
+    state["idx"] = state["idx"] + 1
+    video, meta, prompt = _load_item(state["order"], state["idx"])
+    return state, video, meta, prompt, 3, "", f"⏭️ Skipped. Position: {state['idx']}"
 
 
 # ---------------------------------------------------------------------------
-# Gradio UI
+# Gradio UI — single page (no visibility toggling)
 # ---------------------------------------------------------------------------
 
-THEME = gr.themes.Soft(primary_hue="indigo")
-
-
-with gr.Blocks(theme=THEME, title="MBench-V Annotation") as demo:
+with gr.Blocks(title="MBench-V Annotation", theme=gr.themes.Soft()) as demo:
     gr.Markdown(
         """
         # 🎬 MBench-V Annotation
 
-        Watch each generated video and rate it **1–5** (5 = best). Click **Submit & Next** to save.
-        Your submissions are auto-committed to the dataset repo every 5 minutes.
+        1. Enter a short name (any string — it tags your submissions).
+        2. Click **Start** — a video will appear below.
+        3. Give a score (1–5, 5 = best) and optional note; click **Submit & Next**.
+        4. Submissions auto-sync to the dataset repo every 5 minutes.
         """
     )
 
-    session_state = gr.State(value=None)
+    state = gr.State(value={})
 
-    # ---- Login panel ----
-    with gr.Group(visible=True) as login_panel:
-        with gr.Row():
-            annotator_in = gr.Textbox(
-                label="Annotator name", placeholder="e.g. alice",
-                scale=4, autofocus=True,
+    with gr.Row():
+        annotator_in = gr.Textbox(
+            label="Annotator name",
+            placeholder="e.g. alice",
+            scale=4,
+        )
+        login_btn = gr.Button("Start", variant="primary", scale=1)
+
+    status_md = gr.Markdown("_Not started yet._")
+
+    with gr.Row():
+        with gr.Column(scale=3):
+            video = gr.Video(label="Generated video", autoplay=True, loop=True)
+        with gr.Column(scale=2):
+            meta_md = gr.Markdown()
+            prompt_tb = gr.Textbox(
+                label="Generation prompt",
+                lines=10,
+                max_lines=20,
+                interactive=False,
             )
-            login_btn = gr.Button("Start annotating", variant="primary", scale=1)
-
-    # ---- Annotation panel ----
-    with gr.Group(visible=False) as ann_panel:
-        with gr.Row():
-            with gr.Column(scale=3):
-                video = gr.Video(label="Generated video", autoplay=True, loop=True)
-            with gr.Column(scale=2):
-                meta_md = gr.Markdown()
-                prompt_tb = gr.Textbox(
-                    label="Generation prompt",
-                    lines=10, max_lines=20, interactive=False,
-                )
-            with gr.Column(scale=1):
-                score = gr.Slider(1, 5, value=3, step=1, label="Score (1 worst – 5 best)")
-                note = gr.Textbox(label="Note (optional)", lines=4)
-                submit_btn = gr.Button("✅ Submit & Next", variant="primary")
-                skip_btn = gr.Button("⏭️  Skip")
-
-    status = gr.Markdown("")
-
-    # ---- Wiring ----
+        with gr.Column(scale=1):
+            score = gr.Slider(1, 5, value=3, step=1, label="Score")
+            note = gr.Textbox(label="Note (optional)", lines=4)
+            submit_btn = gr.Button("✅ Submit & Next", variant="primary")
+            skip_btn = gr.Button("⏭️ Skip")
+
     login_btn.click(
         start_session,
-        inputs=[annotator_in, session_state],
-        outputs=[session_state, login_panel, ann_panel, video, meta_md, prompt_tb, status],
+        inputs=[annotator_in, state],
+        outputs=[state, video, meta_md, prompt_tb, status_md],
     )
     annotator_in.submit(
         start_session,
-        inputs=[annotator_in, session_state],
-        outputs=[session_state, login_panel, ann_panel, video, meta_md, prompt_tb, status],
+        inputs=[annotator_in, state],
+        outputs=[state, video, meta_md, prompt_tb, status_md],
     )
     submit_btn.click(
         submit_and_next,
-        inputs=[session_state, score, note],
-        outputs=[session_state, video, meta_md, prompt_tb, score, note, status],
+        inputs=[state, score, note],
+        outputs=[state, video, meta_md, prompt_tb, score, note, status_md],
     )
     skip_btn.click(
         skip_and_next,
-        inputs=[session_state],
-        outputs=[session_state, video, meta_md, prompt_tb, score, note, status],
+        inputs=[state],
+        outputs=[state, video, meta_md, prompt_tb, score, note, status_md],
     )