Fix token_f1 examples to show precision vs recall tradeoff

Old example had precision=recall=0.5 which doesn't teach anything.
New examples show: cautious (high P, low R), overeager (low P, high R),
perfect, and completely wrong.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

Author: abossard

notebooks/01_evaluation.ipynb

@@ -76,28 +76,43 @@
    "metadata": {},
    "outputs": [],
    "source": [
-    "# Metriken werden immer raffinierter\n",
     "from dspy_tasks.calculations import token_f1\n",
     "\n",
-    "f1_val = token_f1([\"apple\", \"banana\"], [\"apple\", \"cherry\"])\n",
-    "\n",
     "print(\"━\" * 50)\n",
     "print(\"📏 Level 1: Exact Match\")\n",
+    "print(\"  Der einfachste Check: stimmt die Antwort exakt?\")\n",
     "print(f\"  'positive' == 'positive' → 1.0 ✅\")\n",
-    "print(f\"  'positive' == 'POSITIVE' → 1.0 ✅ (normalized)\")\n",
     "print(f\"  'positive' == 'negative' → 0.0 ❌\")\n",
     "print()\n",
+    "print(\"━\" * 50)\n",
     "print(\"📏 Level 2: Token F1 (Partial Credit)\")\n",
-    "print(f\"  gold=['apple','banana'], pred=['apple','cherry'] → {f1_val:.3f}\")\n",
+    "print(\"  Was wenn die Antwort TEILWEISE stimmt?\")\n",
+    "print(\"  F1 balanciert Precision (wie viele Treffer waren korrekt?)\")\n",
+    "print(\"  und Recall (wie viele wurden gefunden?)\")\n",
     "print()\n",
-    "print(\"📏 Level 3: Composite (Ticket Routing)\")\n",
-    "print(\"  Priority korrekt (40%) + Kategorie korrekt (35%) + Team korrekt (25%)\")\n",
-    "print(\"  = Gewichtete Spezifikation von 'was am wichtigsten ist'\")\n",
     "\n",
-    "display_insight(\"Der Kern-Insight\",\n",
-    "    \"Deine Metrik-Funktion IST deine Produkt-Spezifikation. \"\n",
-    "    \"Wenn du die Gewichte änderst, änderst du, wofür das System optimiert. \"\n",
-    "    \"Deshalb ist Evaluation die Software der Zukunft.\")"
+    "cases = [\n",
+    "    ([\"apple\", \"banana\", \"cherry\"], [\"apple\"],\n",
+    "     \"Vorsichtig: nur 1 von 3 gefunden, aber der war richtig\"),\n",
+    "    ([\"apple\", \"banana\", \"cherry\"], [\"apple\", \"banana\", \"cherry\", \"grape\", \"melon\"],\n",
+    "     \"Übereifrig: alles gefunden, aber 2 Falsche dabei\"),\n",
+    "    ([\"apple\", \"banana\", \"cherry\"], [\"apple\", \"banana\", \"cherry\"],\n",
+    "     \"Perfekt: genau richtig\"),\n",
+    "    ([\"apple\", \"banana\"], [\"cherry\", \"grape\"],\n",
+    "     \"Komplett daneben: nichts stimmt\"),\n",
+    "]\n",
+    "for gold, pred, desc in cases:\n",
+    "    f1 = token_f1(gold, pred)\n",
+    "    icon = \"✅\" if f1 == 1.0 else \"🟡\" if f1 > 0 else \"❌\"\n",
+    "    print(f\"  {icon} {desc}\")\n",
+    "    print(f\"     Gold: {gold} | Pred: {pred} → F1 = {f1:.2f}\")\n",
+    "\n",
+    "print()\n",
+    "print(\"━\" * 50)\n",
+    "print(\"📏 Level 3: Composite (Ticket Routing)\")\n",
+    "print(\"  Mehrere Kriterien mit Gewichtung:\")\n",
+    "print(\"  Priorität korrekt (40%) + Kategorie (35%) + Team (25%)\")\n",
+    "print(\"  = Gewichtete Spezifikation von 'was am wichtigsten ist'\")\n"
    ]
   },
   {
@@ -305,4 +320,4 @@
  },
  "nbformat": 4,
  "nbformat_minor": 4
-}
+}