Inter font; single Classify view; remove Videos + VLM tabs

README.md

@@ -18,12 +18,9 @@ cell-cycle state (interphase / pre-mitosis / mitosis) directly from short single
 clips — replacing the *classify* stage of the conventional segment → track → classify
 pipeline with a single clip → state model.
 
-This Space serves **pre-computed** results (no heavy model load on the free CPU tier):
-
-1. **Classification gallery** — correct calls + the characteristic failure modes.
-2. **Videos** — whole-FOV predicted-state overlay + a Trackastra-style tracking view.
-3. **Counting** — the same frozen encoder probed for per-state cell counts (honest: beats a null on total, not per-state).
-4. **VLM point-reasoning** — a zero-shot "point at each nucleus then sum" baseline.
+Pick a single-cell clip and the model classifies it into its cell-cycle state — shown
+as **predicted vs actual**. The selector tags each clip ✅ correct / ❌ misclassified so
+you can inspect both. Held-out test-set metrics (n=5,312) are in the collapsible panel.
 
 **Study finding:** in this small-data regime, **data scaling, not model scaling, is the
 binding constraint** — moving the baseline from a small to a larger labelled dataset lifts

app.py

@@ -117,46 +117,29 @@ def _metrics_md() -> str:
     return "\n".join(lines)
 
 
+THEME = gr.themes.Soft(font=[gr.themes.GoogleFont("Inter"), "ui-sans-serif", "system-ui", "sans-serif"])
+
+
 def build():
-    with gr.Blocks(title="Temporal State Prediction", theme=gr.themes.Soft()) as demo:
+    with gr.Blocks(title="Temporal State Prediction", theme=THEME) as demo:
         gr.Markdown(HEADER)
-        with gr.Tabs():
-            with gr.Tab("① Classify a clip"):
-                gr.Markdown("**Select a single-cell clip below** — the model classifies that one clip into its cell-cycle state.")
-                valid = _valid_indices()
-                first = valid[0] if valid else 0
-                img0, md0 = _clip_view(first) if valid else (None, "_no clips found_")
-                with gr.Row():
-                    sel_clip = gr.Image(value=img0, label="selected clip", height=300)
-                    sel_md = gr.Markdown(md0)
-                selector = gr.Gallery(value=_selector_items(), columns=10, height=170,
-                                      object_fit="cover", label="▼ pick a clip (✅ correct · ❌ misclassified)",
-                                      allow_preview=False)
-
-                def _on_select(evt: gr.SelectData):
-                    return _clip_view(evt.index)
-                selector.select(_on_select, inputs=None, outputs=[sel_clip, sel_md])
-
-                with gr.Accordion("Held-out test-set metrics (all 5,312 clips)", open=False):
-                    gr.Markdown(_metrics_md())
-
-            with gr.Tab("② Videos"):
-                gr.Markdown("Whole field-of-view over the held-out sequence.")
-                with gr.Row():
-                    so = ASSETS / "videos" / "state_overlay.mp4"
-                    ti = ASSETS / "videos" / "trackid.mp4"
-                    if so.exists():
-                        gr.Video(str(so), label="Predicted cell-cycle state (blue=interphase, amber=pre-mitosis, red=mitosis)", autoplay=True)
-                    if ti.exists():
-                        gr.Video(str(ti), label="Trackastra-style tracking (colour = track ID)", autoplay=True)
-
-            with gr.Tab("③ VLM point-reasoning (zero-shot)"):
-                gr.Markdown("A frontier VLM prompted to **point at each nucleus while reasoning**, then sum per state — "
-                            "the 'visual primitives' recipe. Zero-shot baseline (no fine-tune); OOD-limited, shown for interpretability.")
-                vlm = ASSETS / "vlm" / "vlm_overlay.png"
-                if vlm.exists():
-                    gr.Image(str(vlm), label="VLM predicted points (○) vs ground-truth centroids (×)")
-                gr.Markdown((ASSETS / "vlm" / "vlm_trace.md").read_text() if (ASSETS / "vlm" / "vlm_trace.md").exists() else "")
+        gr.Markdown("**Select a single-cell clip below** — the model classifies that one clip into its cell-cycle state.")
+        valid = _valid_indices()
+        first = valid[0] if valid else 0
+        img0, md0 = _clip_view(first) if valid else (None, "_no clips found_")
+        with gr.Row():
+            sel_clip = gr.Image(value=img0, label="selected clip", height=300)
+            sel_md = gr.Markdown(md0)
+        selector = gr.Gallery(value=_selector_items(), columns=10, height=170,
+                              object_fit="cover", label="▼ pick a clip (✅ correct · ❌ misclassified)",
+                              allow_preview=False)
+
+        def _on_select(evt: gr.SelectData):
+            return _clip_view(evt.index)
+        selector.select(_on_select, inputs=None, outputs=[sel_clip, sel_md])
+
+        with gr.Accordion("Held-out test-set metrics (all 5,312 clips)", open=False):
+            gr.Markdown(_metrics_md())
 
         gr.Markdown("---\nModels: `DnaRnaProteins/vjepa2-cell-cycle-vit-l`, `DnaRnaProteins/unet-bilstm-cell-cycle-baseline` · "
                     "Data: MICCAI Cell Tracking Challenge (Fluo-N2DL-HeLa). Labels derived from lineage trees (no manual annotation).")

assets/videos/state_overlay.mp4

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:991fc1ed3db265d3b86776566a89f4d911b0969247eaa19b6e9dd169c3fd72cc
-size 22679054

assets/videos/trackid.mp4

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:fe2b552191a1b145602db385554e90040fa542b2c4536ad26ba4982db6cc871e
-size 22132096