diff --git a/examples/predict_tcga_skcm.ipynb b/examples/predict_tcga_skcm.ipynb
index afe339c..7158811 100644
--- a/examples/predict_tcga_skcm.ipynb
+++ b/examples/predict_tcga_skcm.ipynb
@@ -30,15 +30,20 @@
     "tiles with mean pixel above 220, predicts the three genes for each tissue tile, and\n",
     "overlays the predictions on the H&E.\n",
     "\n",
+    "To give a sense of scale, this slide is 99,960 x 65,991 px, which is a 446 x 294\n",
+    "tile grid: 45,751 candidate tiles from the thumbnail pass, of which 36,392 survive\n",
+    "the exact background filter and are scored by the model.\n",
+    "\n",
     "> The TCGA atlas predictions for melanoma were produced with a DeepSpot-M model\n",
     "> finetuned on melanoma spatial transcriptomics. This tutorial uses the base\n",
     "> released model in zero-shot mode, so the values are illustrative and the maps\n",
     "> will look softer than the finetuned atlas.\n",
     "\n",
-    "Install with `pip install deepspotm pyvips matplotlib`. pyvips needs the system\n",
-    "libvips with OpenSlide support. A GPU is recommended. The model weights are gated on\n",
-    "the Hugging Face Hub, so request access and log in first with `huggingface-cli login`.\n",
-    "Note that pyvips must be imported before torch."
+    "Install with `pip install \"deepspotm[wsi]\" matplotlib`, matching the `wsi` extra\n",
+    "documented in the README. pyvips needs the system libvips with OpenSlide support.\n",
+    "A GPU is recommended. The model weights are gated on the Hugging Face Hub, so\n",
+    "request access and log in first with `huggingface-cli login`. Note that pyvips\n",
+    "must be imported before torch."
    ]
   },
   {
@@ -107,8 +112,13 @@
     "    print(\"downloading ~1 GB slide from GDC ...\")\n",
     "    urllib.request.urlretrieve(url, SLIDE_PATH)\n",
     "\n",
+    "# Hash in 8 MB chunks; reading the whole ~1 GB slide into memory to checksum it\n",
+    "# would spike RSS by more than the model itself uses.\n",
+    "digest = hashlib.md5()\n",
     "with open(SLIDE_PATH, \"rb\") as fh:\n",
-    "    md5 = hashlib.md5(fh.read()).hexdigest()\n",
+    "    for chunk in iter(lambda: fh.read(8 << 20), b\"\"):\n",
+    "        digest.update(chunk)\n",
+    "md5 = digest.hexdigest()\n",
     "assert md5 == SLIDE_MD5, f\"checksum mismatch: {md5}\"\n",
     "print(\"slide ready:\", SLIDE_PATH)"
    ]