Fix ELECTRA examples with locally-verified sentences that trigger detection

Part 4: Replaced doctor/watched examples (no detection) with sentences
verified locally against electra-small-discriminator:
- 'third quarter' → 'third half': half flagged at 0.586 (semantic violation)
- 'mysterious case' → 'mysterious mystery': flagged at 0.570 (redundancy)
- Visualization: 'painted' → 'created' flags 'landscape' at 0.526
Updated discussion questions to highlight context propagation.

Author: jeremymanning

slides/week6/bert_variants_demo.ipynb

@@ -340,17 +340,18 @@
         "        print(f\"{token:<12} | {prob:<10.4f} | {status}\")\n",
         "\n",
         "# Example 1: A natural sentence (all tokens should be marked REAL)\n",
-        "detect_fake_tokens(\"The doctor examined the patient carefully.\")\n",
+        "detect_fake_tokens(\"The company announced record profits for the third quarter.\")\n",
         "\n",
         "print(\"\\n\" + \"=\"*40 + \"\\n\")\n",
         "\n",
-        "# Example 2: Same sentence but 'examined' → 'watched' (plausible but wrong in context)\n",
-        "detect_fake_tokens(\"The doctor watched the patient carefully.\")\n",
+        "# Example 2: 'quarter' replaced with 'half' — grammatical but semantically wrong\n",
+        "# (companies report in quarters, not halves)\n",
+        "detect_fake_tokens(\"The company announced record profits for the third half.\")\n",
         "\n",
         "print(\"\\n\" + \"=\"*40 + \"\\n\")\n",
         "\n",
-        "# Example 3: 'doctor' → 'musician' (semantically odd with 'patient')\n",
-        "detect_fake_tokens(\"The musician examined the patient carefully.\")"
+        "# Example 3: 'case' replaced with 'mystery' — creates redundancy with 'mysterious'\n",
+        "detect_fake_tokens(\"The detective solved the mysterious mystery in just two days.\")"
       ]
     },
     {
@@ -360,8 +361,9 @@
       "outputs": [],
       "source": [
         "# Let's visualize the discriminator's confidence\n",
-        "# 'published' has been swapped in for 'presented' — plausible but detectable\n",
-        "sentence = \"The researcher published her findings at the annual conference.\"\n",
+        "# 'painted' has been swapped to 'created' — plausible synonym, but ELECTRA\n",
+        "# flags 'landscape' as suspicious (you 'paint' a landscape, not 'create' one)\n",
+        "sentence = \"The artist created a beautiful landscape of the countryside.\"\n",
         "inputs = tokenizer(sentence, return_tensors=\"pt\")\n",
         "tokens = tokenizer.convert_ids_to_tokens(inputs[\"input_ids\"][0])[1:-1] # Remove CLS/SEP\n",
         "\n",
@@ -387,8 +389,9 @@
       "source": [
         "### \ud83d\udca1 Discussion\n",
         "\n",
-        "- In Example 2, did ELECTRA flag \"watched\" as suspicious? Doctors *examine* patients — \"watched\" is grammatical but semantically odd in a clinical context.\n",
-        "- In Example 3, did swapping \"doctor\" → \"musician\" get detected? Why might that be easier or harder to detect than swapping the verb?\n",
+        "- In Example 2, did ELECTRA flag \"half\" as fake? Why is \"third half\" suspicious even though it's grammatically correct?\n",
+        "- In Example 3, the discriminator flags \"mysterious\" rather than \"mystery\" — why might it flag the *context* around a fake token rather than the fake token itself?\n",
+        "- In the bar chart, ELECTRA flags \"landscape\" as fake even though it's the *original* word. The swap was \"painted\" → \"created\". Why does suspicion propagate to neighboring tokens?\n",
         "- How is this task different from BERT's masked language modeling? Why might it be more efficient? (Hint: how many tokens does each approach learn from per sentence?)\n",
         "- Try replacing a word with a close synonym (e.g., \"quickly\" → \"rapidly\"). Does ELECTRA flag it?"
       ]