Upload post_training_eval.json with huggingface_hub

post_training_eval.json

@@ -4,107 +4,107 @@
       "prompt": "Write a 200-word blog post about the benefits of remote work.",
       "variants": {
         "ai_original": {
-          "text_length": 188,
-          "gpt2_perplexity": 17.9,
-          "gpt2_burstiness": 0.6834,
-          "fast_detectgpt": 0.0002,
-          "gptneo_perplexity": 27.8,
-          "char_distribution": 0.7382,
-          "stylometric": 0.325,
+          "text_length": 185,
+          "gpt2_perplexity": 19.4,
+          "gpt2_burstiness": 1.0828,
+          "fast_detectgpt": 0.0001,
+          "gptneo_perplexity": 24.6,
+          "char_distribution": 0.7454,
+          "stylometric": 0.3,
           "stylo_details": {
             "sentence_count": 12,
-            "sentence_length_mean": 15.7,
-            "sentence_length_std": 5.3,
+            "sentence_length_mean": 15.4,
+            "sentence_length_std": 6.5,
             "total_words": 189,
-            "unique_words": 130,
-            "lexical_diversity": 0.688,
-            "hapax_ratio": 0.566,
-            "avg_word_length": 5.3,
-            "word_length_std": 3.1,
-            "transition_rate": 0.5,
+            "unique_words": 136,
+            "lexical_diversity": 0.72,
+            "hapax_ratio": 0.598,
+            "avg_word_length": 5.6,
+            "word_length_std": 3.2,
+            "transition_rate": 1.1,
             "passive_rate": 0.0,
             "formulaic_rate": 0.0,
-            "punct_ratio": 0.0244
+            "punct_ratio": 0.0236
           },
-          "ensemble_score": 0.4928,
+          "ensemble_score": 0.4689,
           "ensemble_class": "Uncertain"
         },
         "humanized_std": {
-          "text_length": 213,
-          "gpt2_perplexity": 19.6,
-          "gpt2_burstiness": 0.6646,
+          "text_length": 215,
+          "gpt2_perplexity": 19.4,
+          "gpt2_burstiness": 1.1177,
           "fast_detectgpt": 0.0001,
-          "gptneo_perplexity": 26.2,
-          "char_distribution": 0.7464,
-          "stylometric": 0.325,
+          "gptneo_perplexity": 24.3,
+          "char_distribution": 0.7692,
+          "stylometric": 0.3167,
           "stylo_details": {
-            "sentence_count": 14,
-            "sentence_length_mean": 15.2,
-            "sentence_length_std": 5.5,
-            "total_words": 218,
-            "unique_words": 146,
-            "lexical_diversity": 0.67,
-            "hapax_ratio": 0.546,
-            "avg_word_length": 5.2,
-            "word_length_std": 3.0,
-            "transition_rate": 0.5,
+            "sentence_count": 13,
+            "sentence_length_mean": 16.5,
+            "sentence_length_std": 7.4,
+            "total_words": 219,
+            "unique_words": 148,
+            "lexical_diversity": 0.676,
+            "hapax_ratio": 0.553,
+            "avg_word_length": 5.4,
+            "word_length_std": 3.1,
+            "transition_rate": 0.9,
             "passive_rate": 0.0,
             "formulaic_rate": 0.0,
-            "punct_ratio": 0.0238
+            "punct_ratio": 0.0231
           },
-          "ensemble_score": 0.4483,
+          "ensemble_score": 0.4744,
           "ensemble_class": "Uncertain"
         },
         "humanized_agg": {
           "text_length": 213,
-          "gpt2_perplexity": 19.6,
-          "gpt2_burstiness": 0.6582,
+          "gpt2_perplexity": 19.5,
+          "gpt2_burstiness": 1.1025,
           "fast_detectgpt": 0.0001,
-          "gptneo_perplexity": 26.7,
-          "char_distribution": 0.749,
-          "stylometric": 0.325,
+          "gptneo_perplexity": 24.2,
+          "char_distribution": 0.7735,
+          "stylometric": 0.3167,
           "stylo_details": {
             "sentence_count": 14,
-            "sentence_length_mean": 15.2,
-            "sentence_length_std": 5.5,
-            "total_words": 218,
-            "unique_words": 147,
-            "lexical_diversity": 0.674,
-            "hapax_ratio": 0.55,
-            "avg_word_length": 5.2,
+            "sentence_length_mean": 15.3,
+            "sentence_length_std": 6.3,
+            "total_words": 217,
+            "unique_words": 148,
+            "lexical_diversity": 0.682,
+            "hapax_ratio": 0.558,
+            "avg_word_length": 5.4,
             "word_length_std": 3.1,
-            "transition_rate": 0.5,
+            "transition_rate": 0.9,
             "passive_rate": 0.0,
             "formulaic_rate": 0.0,
-            "punct_ratio": 0.0238
+            "punct_ratio": 0.0247
           },
-          "ensemble_score": 0.4483,
+          "ensemble_score": 0.4744,
           "ensemble_class": "Uncertain"
         },
         "humanized_scratch": {
-          "text_length": 210,
-          "gpt2_perplexity": 18.4,
-          "gpt2_burstiness": 2.139,
-          "fast_detectgpt": 0.0002,
-          "gptneo_perplexity": 24.7,
-          "char_distribution": 0.7816,
-          "stylometric": 0.35,
+          "text_length": 217,
+          "gpt2_perplexity": 19.8,
+          "gpt2_burstiness": 2.2795,
+          "fast_detectgpt": 0.0001,
+          "gptneo_perplexity": 25.1,
+          "char_distribution": 0.7372,
+          "stylometric": 0.2917,
           "stylo_details": {
-            "sentence_count": 14,
-            "sentence_length_mean": 15.0,
-            "sentence_length_std": 6.8,
-            "total_words": 218,
-            "unique_words": 144,
-            "lexical_diversity": 0.661,
-            "hapax_ratio": 0.514,
-            "avg_word_length": 4.3,
+            "sentence_count": 11,
+            "sentence_length_mean": 19.7,
+            "sentence_length_std": 8.9,
+            "total_words": 223,
+            "unique_words": 151,
+            "lexical_diversity": 0.677,
+            "hapax_ratio": 0.525,
+            "avg_word_length": 4.4,
             "word_length_std": 2.2,
-            "transition_rate": 0.5,
-            "passive_rate": 0.5,
+            "transition_rate": 0.0,
+            "passive_rate": 0.0,
             "formulaic_rate": 0.0,
-            "punct_ratio": 0.0246
+            "punct_ratio": 0.0229
           },
-          "ensemble_score": 0.4856,
+          "ensemble_score": 0.4366,
           "ensemble_class": "Uncertain"
         }
       }
@@ -113,107 +113,107 @@
       "prompt": "Explain quantum computing in simple terms, around 150 words.",
       "variants": {
         "ai_original": {
-          "text_length": 148,
-          "gpt2_perplexity": 17.9,
-          "gpt2_burstiness": 1.0057,
+          "text_length": 143,
+          "gpt2_perplexity": 18.0,
+          "gpt2_burstiness": 0.4443,
           "fast_detectgpt": 0.0002,
-          "gptneo_perplexity": 30.3,
-          "char_distribution": 0.5996,
-          "stylometric": 0.2333,
+          "gptneo_perplexity": 28.0,
+          "char_distribution": 0.5755,
+          "stylometric": 0.2667,
           "stylo_details": {
             "sentence_count": 8,
-            "sentence_length_mean": 18.6,
-            "sentence_length_std": 9.3,
-            "total_words": 156,
-            "unique_words": 110,
-            "lexical_diversity": 0.705,
-            "hapax_ratio": 0.526,
+            "sentence_length_mean": 18.0,
+            "sentence_length_std": 6.1,
+            "total_words": 149,
+            "unique_words": 102,
+            "lexical_diversity": 0.685,
+            "hapax_ratio": 0.51,
             "avg_word_length": 4.6,
             "word_length_std": 2.8,
             "transition_rate": 0.0,
             "passive_rate": 0.0,
             "formulaic_rate": 0.0,
-            "punct_ratio": 0.0301
+            "punct_ratio": 0.031
           },
-          "ensemble_score": 0.3929,
-          "ensemble_class": "Human"
+          "ensemble_score": 0.4048,
+          "ensemble_class": "Uncertain"
         },
         "humanized_std": {
-          "text_length": 200,
-          "gpt2_perplexity": 16.7,
-          "gpt2_burstiness": 0.9667,
-          "fast_detectgpt": 0.0002,
-          "gptneo_perplexity": 24.9,
-          "char_distribution": 0.6324,
-          "stylometric": 0.325,
+          "text_length": 197,
+          "gpt2_perplexity": 15.0,
+          "gpt2_burstiness": 1.127,
+          "fast_detectgpt": 0.0003,
+          "gptneo_perplexity": 23.0,
+          "char_distribution": 0.6265,
+          "stylometric": 0.2917,
           "stylo_details": {
-            "sentence_count": 11,
-            "sentence_length_mean": 18.4,
-            "sentence_length_std": 8.8,
-            "total_words": 210,
-            "unique_words": 124,
-            "lexical_diversity": 0.59,
-            "hapax_ratio": 0.367,
-            "avg_word_length": 4.6,
+            "sentence_count": 12,
+            "sentence_length_mean": 16.7,
+            "sentence_length_std": 6.4,
+            "total_words": 205,
+            "unique_words": 127,
+            "lexical_diversity": 0.62,
+            "hapax_ratio": 0.424,
+            "avg_word_length": 4.7,
             "word_length_std": 2.7,
             "transition_rate": 0.0,
             "passive_rate": 0.0,
             "formulaic_rate": 0.0,
-            "punct_ratio": 0.0308
+            "punct_ratio": 0.0328
           },
-          "ensemble_score": 0.4563,
+          "ensemble_score": 0.4449,
           "ensemble_class": "Uncertain"
         },
         "humanized_agg": {
-          "text_length": 195,
-          "gpt2_perplexity": 20.7,
-          "gpt2_burstiness": 0.9057,
-          "fast_detectgpt": 0.0001,
-          "gptneo_perplexity": 30.2,
-          "char_distribution": 0.6447,
-          "stylometric": 0.2583,
+          "text_length": 196,
+          "gpt2_perplexity": 14.7,
+          "gpt2_burstiness": 0.5508,
+          "fast_detectgpt": 0.0003,
+          "gptneo_perplexity": 21.9,
+          "char_distribution": 0.6583,
+          "stylometric": 0.325,
           "stylo_details": {
             "sentence_count": 12,
             "sentence_length_mean": 16.4,
-            "sentence_length_std": 9.2,
-            "total_words": 207,
-            "unique_words": 134,
-            "lexical_diversity": 0.647,
-            "hapax_ratio": 0.444,
+            "sentence_length_std": 6.3,
+            "total_words": 205,
+            "unique_words": 125,
+            "lexical_diversity": 0.61,
+            "hapax_ratio": 0.415,
             "avg_word_length": 4.6,
             "word_length_std": 2.7,
             "transition_rate": 0.0,
             "passive_rate": 0.0,
             "formulaic_rate": 0.0,
-            "punct_ratio": 0.0311
+            "punct_ratio": 0.0335
           },
-          "ensemble_score": 0.3494,
-          "ensemble_class": "Human"
+          "ensemble_score": 0.4563,
+          "ensemble_class": "Uncertain"
         },
         "humanized_scratch": {
-          "text_length": 198,
-          "gpt2_perplexity": 15.9,
-          "gpt2_burstiness": 0.5911,
-          "fast_detectgpt": 0.0003,
-          "gptneo_perplexity": 25.8,
-          "char_distribution": 0.6939,
-          "stylometric": 0.375,
+          "text_length": 186,
+          "gpt2_perplexity": 17.1,
+          "gpt2_burstiness": 0.4344,
+          "fast_detectgpt": 0.0002,
+          "gptneo_perplexity": 26.4,
+          "char_distribution": 0.6822,
+          "stylometric": 0.2917,
           "stylo_details": {
             "sentence_count": 11,
-            "sentence_length_mean": 18.1,
-            "sentence_length_std": 4.7,
-            "total_words": 212,
-            "unique_words": 132,
-            "lexical_diversity": 0.623,
-            "hapax_ratio": 0.453,
-            "avg_word_length": 4.2,
-            "word_length_std": 2.5,
+            "sentence_length_mean": 17.0,
+            "sentence_length_std": 6.4,
+            "total_words": 198,
+            "unique_words": 131,
+            "lexical_diversity": 0.662,
+            "hapax_ratio": 0.505,
+            "avg_word_length": 4.5,
+            "word_length_std": 2.6,
             "transition_rate": 0.5,
             "passive_rate": 0.0,
             "formulaic_rate": 0.0,
-            "punct_ratio": 0.0302
+            "punct_ratio": 0.0312
           },
-          "ensemble_score": 0.4435,
+          "ensemble_score": 0.4137,
           "ensemble_class": "Uncertain"
         }
       }
@@ -222,115 +222,115 @@
       "prompt": "Write a professional email declining a job offer, about 100 words.",
       "variants": {
         "ai_original": {
-          "text_length": 103,
-          "gpt2_perplexity": 8.8,
-          "gpt2_burstiness": 0.5193,
-          "fast_detectgpt": 0.0015,
-          "gptneo_perplexity": 13.2,
-          "char_distribution": 0.5868,
-          "stylometric": 0.2667,
+          "text_length": 102,
+          "gpt2_perplexity": 7.4,
+          "gpt2_burstiness": 0.5157,
+          "fast_detectgpt": 0.0025,
+          "gptneo_perplexity": 11.9,
+          "char_distribution": 0.6029,
+          "stylometric": 0.2917,
           "stylo_details": {
             "sentence_count": 6,
-            "sentence_length_mean": 17.2,
-            "sentence_length_std": 7.1,
+            "sentence_length_mean": 17.0,
+            "sentence_length_std": 8.0,
             "total_words": 101,
             "unique_words": 68,
             "lexical_diversity": 0.673,
-            "hapax_ratio": 0.515,
-            "avg_word_length": 4.9,
-            "word_length_std": 2.6,
+            "hapax_ratio": 0.505,
+            "avg_word_length": 4.7,
+            "word_length_std": 2.5,
             "transition_rate": 0.0,
             "passive_rate": 0.0,
             "formulaic_rate": 0.0,
-            "punct_ratio": 0.0397
+            "punct_ratio": 0.0429
           },
-          "ensemble_score": 0.5375,
+          "ensemble_score": 0.5453,
           "ensemble_class": "AI"
         },
         "humanized_std": {
-          "text_length": 198,
-          "gpt2_perplexity": 8.1,
-          "gpt2_burstiness": 0.4646,
-          "fast_detectgpt": 0.0019,
-          "gptneo_perplexity": 10.1,
-          "char_distribution": 0.6681,
-          "stylometric": 0.35,
+          "text_length": 197,
+          "gpt2_perplexity": 8.7,
+          "gpt2_burstiness": 0.6917,
+          "fast_detectgpt": 0.0015,
+          "gptneo_perplexity": 10.6,
+          "char_distribution": 0.6639,
+          "stylometric": 0.3833,
           "stylo_details": {
-            "sentence_count": 11,
-            "sentence_length_mean": 18.0,
-            "sentence_length_std": 7.6,
-            "total_words": 197,
-            "unique_words": 110,
-            "lexical_diversity": 0.558,
-            "hapax_ratio": 0.376,
-            "avg_word_length": 4.5,
-            "word_length_std": 2.4,
+            "sentence_count": 12,
+            "sentence_length_mean": 16.4,
+            "sentence_length_std": 6.0,
+            "total_words": 199,
+            "unique_words": 106,
+            "lexical_diversity": 0.533,
+            "hapax_ratio": 0.337,
+            "avg_word_length": 4.4,
+            "word_length_std": 2.3,
             "transition_rate": 0.0,
             "passive_rate": 0.0,
             "formulaic_rate": 0.0,
-            "punct_ratio": 0.0429
+            "punct_ratio": 0.037
           },
-          "ensemble_score": 0.5635,
+          "ensemble_score": 0.574,
           "ensemble_class": "AI"
         },
         "humanized_agg": {
-          "text_length": 191,
-          "gpt2_perplexity": 14.5,
-          "gpt2_burstiness": 0.8113,
-          "fast_detectgpt": 0.0003,
-          "gptneo_perplexity": 18.0,
-          "char_distribution": 0.743,
-          "stylometric": 0.2833,
+          "text_length": 196,
+          "gpt2_perplexity": 9.1,
+          "gpt2_burstiness": 0.668,
+          "fast_detectgpt": 0.0013,
+          "gptneo_perplexity": 11.6,
+          "char_distribution": 0.6982,
+          "stylometric": 0.4167,
           "stylo_details": {
-            "sentence_count": 9,
-            "sentence_length_mean": 21.2,
-            "sentence_length_std": 15.0,
-            "total_words": 198,
-            "unique_words": 120,
-            "lexical_diversity": 0.606,
-            "hapax_ratio": 0.47,
-            "avg_word_length": 4.1,
-            "word_length_std": 2.2,
-            "transition_rate": 0.5,
+            "sentence_count": 12,
+            "sentence_length_mean": 16.3,
+            "sentence_length_std": 5.8,
+            "total_words": 202,
+            "unique_words": 109,
+            "lexical_diversity": 0.54,
+            "hapax_ratio": 0.347,
+            "avg_word_length": 4.4,
+            "word_length_std": 2.3,
+            "transition_rate": 0.0,
             "passive_rate": 0.0,
             "formulaic_rate": 0.0,
-            "punct_ratio": 0.0345
+            "punct_ratio": 0.0352
           },
-          "ensemble_score": 0.5053,
+          "ensemble_score": 0.5844,
           "ensemble_class": "AI"
         },
         "humanized_scratch": {
-          "text_length": 117,
-          "gpt2_perplexity": 11.2,
-          "gpt2_burstiness": 0.5365,
-          "fast_detectgpt": 0.0007,
+          "text_length": 119,
+          "gpt2_perplexity": 9.9,
+          "gpt2_burstiness": 0.6588,
+          "fast_detectgpt": 0.001,
           "gptneo_perplexity": 14.7,
-          "char_distribution": 0.6893,
-          "stylometric": 0.2917,
+          "char_distribution": 0.6348,
+          "stylometric": 0.3167,
           "stylo_details": {
             "sentence_count": 7,
-            "sentence_length_mean": 16.7,
-            "sentence_length_std": 6.4,
-            "total_words": 118,
-            "unique_words": 78,
-            "lexical_diversity": 0.661,
-            "hapax_ratio": 0.5,
-            "avg_word_length": 4.4,
-            "word_length_std": 2.3,
+            "sentence_length_mean": 17.0,
+            "sentence_length_std": 6.2,
+            "total_words": 120,
+            "unique_words": 83,
+            "lexical_diversity": 0.692,
+            "hapax_ratio": 0.533,
+            "avg_word_length": 4.1,
+            "word_length_std": 2.1,
             "transition_rate": 0.0,
             "passive_rate": 0.0,
             "formulaic_rate": 0.0,
-            "punct_ratio": 0.0405
+            "punct_ratio": 0.0416
           },
-          "ensemble_score": 0.5453,
+          "ensemble_score": 0.5531,
           "ensemble_class": "AI"
         }
       }
     }
   ],
   "summary": {
-    "ai_mean": 0.4744,
-    "humanized_mean": 0.4717,
-    "improvement": 0.0027
+    "ai_mean": 0.473,
+    "humanized_mean": 0.4902,
+    "improvement": -0.0172
   }
 }