Upload post_training_eval.json with huggingface_hub

post_training_eval.json

@@ -4,108 +4,108 @@
       "prompt": "Write a 200-word blog post about the benefits of remote work.",
       "variants": {
         "ai_original": {
-          "text_length": 185,
-          "gpt2_perplexity": 19.4,
-          "gpt2_burstiness": 1.0828,
+          "text_length": 184,
+          "gpt2_perplexity": 19.7,
+          "gpt2_burstiness": 0.6946,
           "fast_detectgpt": 0.0001,
-          "gptneo_perplexity": 24.6,
-          "char_distribution": 0.7454,
-          "stylometric": 0.3,
+          "gptneo_perplexity": 26.1,
+          "char_distribution": 0.7568,
+          "stylometric": 0.3083,
           "stylo_details": {
-            "sentence_count": 12,
-            "sentence_length_mean": 15.4,
-            "sentence_length_std": 6.5,
-            "total_words": 189,
-            "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,
+            "sentence_count": 11,
+            "sentence_length_mean": 16.7,
+            "sentence_length_std": 5.6,
+            "total_words": 187,
+            "unique_words": 133,
+            "lexical_diversity": 0.711,
+            "hapax_ratio": 0.583,
+            "avg_word_length": 5.4,
+            "word_length_std": 3.0,
+            "transition_rate": 0.5,
             "passive_rate": 0.0,
             "formulaic_rate": 0.0,
-            "punct_ratio": 0.0236
+            "punct_ratio": 0.0229
           },
-          "ensemble_score": 0.4689,
+          "ensemble_score": 0.4424,
           "ensemble_class": "Uncertain"
         },
         "humanized_std": {
-          "text_length": 215,
-          "gpt2_perplexity": 19.4,
-          "gpt2_burstiness": 1.1177,
-          "fast_detectgpt": 0.0001,
-          "gptneo_perplexity": 24.3,
-          "char_distribution": 0.7692,
+          "text_length": 214,
+          "gpt2_perplexity": 17.8,
+          "gpt2_burstiness": 0.7141,
+          "fast_detectgpt": 0.0002,
+          "gptneo_perplexity": 22.5,
+          "char_distribution": 0.7951,
           "stylometric": 0.3167,
           "stylo_details": {
-            "sentence_count": 13,
-            "sentence_length_mean": 16.5,
-            "sentence_length_std": 7.4,
+            "sentence_count": 12,
+            "sentence_length_mean": 17.8,
+            "sentence_length_std": 6.5,
             "total_words": 219,
-            "unique_words": 148,
-            "lexical_diversity": 0.676,
-            "hapax_ratio": 0.553,
-            "avg_word_length": 5.4,
-            "word_length_std": 3.1,
+            "unique_words": 149,
+            "lexical_diversity": 0.68,
+            "hapax_ratio": 0.548,
+            "avg_word_length": 5.2,
+            "word_length_std": 2.9,
             "transition_rate": 0.9,
             "passive_rate": 0.0,
             "formulaic_rate": 0.0,
-            "punct_ratio": 0.0231
+            "punct_ratio": 0.0224
           },
-          "ensemble_score": 0.4744,
-          "ensemble_class": "Uncertain"
+          "ensemble_score": 0.516,
+          "ensemble_class": "AI"
         },
         "humanized_agg": {
           "text_length": 213,
-          "gpt2_perplexity": 19.5,
-          "gpt2_burstiness": 1.1025,
+          "gpt2_perplexity": 19.0,
+          "gpt2_burstiness": 0.7387,
           "fast_detectgpt": 0.0001,
-          "gptneo_perplexity": 24.2,
-          "char_distribution": 0.7735,
+          "gptneo_perplexity": 25.4,
+          "char_distribution": 0.8057,
           "stylometric": 0.3167,
           "stylo_details": {
-            "sentence_count": 14,
-            "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,
+            "sentence_count": 12,
+            "sentence_length_mean": 17.8,
+            "sentence_length_std": 6.4,
+            "total_words": 218,
+            "unique_words": 149,
+            "lexical_diversity": 0.683,
+            "hapax_ratio": 0.555,
+            "avg_word_length": 5.2,
+            "word_length_std": 2.9,
             "transition_rate": 0.9,
             "passive_rate": 0.0,
             "formulaic_rate": 0.0,
-            "punct_ratio": 0.0247
+            "punct_ratio": 0.0233
           },
-          "ensemble_score": 0.4744,
+          "ensemble_score": 0.4453,
           "ensemble_class": "Uncertain"
         },
         "humanized_scratch": {
           "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,
+          "gpt2_perplexity": 15.9,
+          "gpt2_burstiness": 2.0848,
+          "fast_detectgpt": 0.0002,
+          "gptneo_perplexity": 21.0,
+          "char_distribution": 0.7613,
+          "stylometric": 0.3417,
           "stylo_details": {
-            "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": 0.0,
+            "sentence_count": 12,
+            "sentence_length_mean": 18.1,
+            "sentence_length_std": 6.6,
+            "total_words": 225,
+            "unique_words": 148,
+            "lexical_diversity": 0.658,
+            "hapax_ratio": 0.511,
+            "avg_word_length": 4.2,
+            "word_length_std": 2.1,
+            "transition_rate": 0.4,
             "passive_rate": 0.0,
             "formulaic_rate": 0.0,
-            "punct_ratio": 0.0229
+            "punct_ratio": 0.0219
           },
-          "ensemble_score": 0.4366,
-          "ensemble_class": "Uncertain"
+          "ensemble_score": 0.5239,
+          "ensemble_class": "AI"
         }
       }
     },
@@ -113,105 +113,105 @@
       "prompt": "Explain quantum computing in simple terms, around 150 words.",
       "variants": {
         "ai_original": {
-          "text_length": 143,
-          "gpt2_perplexity": 18.0,
-          "gpt2_burstiness": 0.4443,
-          "fast_detectgpt": 0.0002,
-          "gptneo_perplexity": 28.0,
-          "char_distribution": 0.5755,
+          "text_length": 148,
+          "gpt2_perplexity": 19.4,
+          "gpt2_burstiness": 0.7225,
+          "fast_detectgpt": 0.0001,
+          "gptneo_perplexity": 30.1,
+          "char_distribution": 0.5882,
           "stylometric": 0.2667,
           "stylo_details": {
             "sentence_count": 8,
-            "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,
+            "sentence_length_mean": 18.6,
+            "sentence_length_std": 7.9,
+            "total_words": 151,
+            "unique_words": 109,
+            "lexical_diversity": 0.722,
+            "hapax_ratio": 0.55,
+            "avg_word_length": 4.7,
             "word_length_std": 2.8,
             "transition_rate": 0.0,
             "passive_rate": 0.0,
             "formulaic_rate": 0.0,
-            "punct_ratio": 0.031
+            "punct_ratio": 0.0327
           },
-          "ensemble_score": 0.4048,
-          "ensemble_class": "Uncertain"
+          "ensemble_score": 0.3525,
+          "ensemble_class": "Human"
         },
         "humanized_std": {
-          "text_length": 197,
-          "gpt2_perplexity": 15.0,
-          "gpt2_burstiness": 1.127,
+          "text_length": 202,
+          "gpt2_perplexity": 15.5,
+          "gpt2_burstiness": 0.7687,
           "fast_detectgpt": 0.0003,
-          "gptneo_perplexity": 23.0,
-          "char_distribution": 0.6265,
-          "stylometric": 0.2917,
+          "gptneo_perplexity": 22.2,
+          "char_distribution": 0.6314,
+          "stylometric": 0.325,
           "stylo_details": {
-            "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,
+            "sentence_count": 11,
+            "sentence_length_mean": 18.5,
+            "sentence_length_std": 7.7,
+            "total_words": 206,
+            "unique_words": 126,
+            "lexical_diversity": 0.612,
+            "hapax_ratio": 0.398,
+            "avg_word_length": 4.6,
             "word_length_std": 2.7,
             "transition_rate": 0.0,
             "passive_rate": 0.0,
             "formulaic_rate": 0.0,
-            "punct_ratio": 0.0328
+            "punct_ratio": 0.0368
           },
-          "ensemble_score": 0.4449,
+          "ensemble_score": 0.4563,
           "ensemble_class": "Uncertain"
         },
         "humanized_agg": {
-          "text_length": 196,
-          "gpt2_perplexity": 14.7,
-          "gpt2_burstiness": 0.5508,
+          "text_length": 204,
+          "gpt2_perplexity": 15.6,
+          "gpt2_burstiness": 0.798,
           "fast_detectgpt": 0.0003,
-          "gptneo_perplexity": 21.9,
-          "char_distribution": 0.6583,
+          "gptneo_perplexity": 22.0,
+          "char_distribution": 0.6215,
           "stylometric": 0.325,
           "stylo_details": {
             "sentence_count": 12,
-            "sentence_length_mean": 16.4,
-            "sentence_length_std": 6.3,
-            "total_words": 205,
-            "unique_words": 125,
-            "lexical_diversity": 0.61,
-            "hapax_ratio": 0.415,
+            "sentence_length_mean": 17.2,
+            "sentence_length_std": 7.9,
+            "total_words": 208,
+            "unique_words": 126,
+            "lexical_diversity": 0.606,
+            "hapax_ratio": 0.394,
             "avg_word_length": 4.6,
             "word_length_std": 2.7,
             "transition_rate": 0.0,
             "passive_rate": 0.0,
             "formulaic_rate": 0.0,
-            "punct_ratio": 0.0335
+            "punct_ratio": 0.035
           },
           "ensemble_score": 0.4563,
           "ensemble_class": "Uncertain"
         },
         "humanized_scratch": {
-          "text_length": 186,
-          "gpt2_perplexity": 17.1,
-          "gpt2_burstiness": 0.4344,
+          "text_length": 199,
+          "gpt2_perplexity": 17.8,
+          "gpt2_burstiness": 0.8459,
           "fast_detectgpt": 0.0002,
           "gptneo_perplexity": 26.4,
-          "char_distribution": 0.6822,
+          "char_distribution": 0.6709,
           "stylometric": 0.2917,
           "stylo_details": {
-            "sentence_count": 11,
-            "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,
+            "sentence_count": 10,
+            "sentence_length_mean": 20.0,
+            "sentence_length_std": 6.2,
+            "total_words": 211,
+            "unique_words": 142,
+            "lexical_diversity": 0.673,
+            "hapax_ratio": 0.517,
+            "avg_word_length": 4.4,
             "word_length_std": 2.6,
             "transition_rate": 0.5,
             "passive_rate": 0.0,
             "formulaic_rate": 0.0,
-            "punct_ratio": 0.0312
+            "punct_ratio": 0.0314
           },
           "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": 102,
-          "gpt2_perplexity": 7.4,
-          "gpt2_burstiness": 0.5157,
-          "fast_detectgpt": 0.0025,
-          "gptneo_perplexity": 11.9,
-          "char_distribution": 0.6029,
+          "text_length": 98,
+          "gpt2_perplexity": 8.0,
+          "gpt2_burstiness": 0.6268,
+          "fast_detectgpt": 0.0019,
+          "gptneo_perplexity": 13.6,
+          "char_distribution": 0.5733,
           "stylometric": 0.2917,
           "stylo_details": {
             "sentence_count": 6,
-            "sentence_length_mean": 17.0,
-            "sentence_length_std": 8.0,
-            "total_words": 101,
+            "sentence_length_mean": 16.3,
+            "sentence_length_std": 7.6,
+            "total_words": 97,
             "unique_words": 68,
-            "lexical_diversity": 0.673,
-            "hapax_ratio": 0.505,
-            "avg_word_length": 4.7,
+            "lexical_diversity": 0.701,
+            "hapax_ratio": 0.546,
+            "avg_word_length": 4.9,
             "word_length_std": 2.5,
             "transition_rate": 0.0,
             "passive_rate": 0.0,
             "formulaic_rate": 0.0,
-            "punct_ratio": 0.0429
+            "punct_ratio": 0.043
           },
           "ensemble_score": 0.5453,
           "ensemble_class": "AI"
         },
         "humanized_std": {
-          "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,
+          "text_length": 195,
+          "gpt2_perplexity": 8.9,
+          "gpt2_burstiness": 0.6271,
+          "fast_detectgpt": 0.0014,
+          "gptneo_perplexity": 12.1,
+          "char_distribution": 0.6574,
+          "stylometric": 0.4167,
           "stylo_details": {
             "sentence_count": 12,
-            "sentence_length_mean": 16.4,
-            "sentence_length_std": 6.0,
+            "sentence_length_mean": 16.2,
+            "sentence_length_std": 5.8,
             "total_words": 199,
-            "unique_words": 106,
-            "lexical_diversity": 0.533,
-            "hapax_ratio": 0.337,
-            "avg_word_length": 4.4,
-            "word_length_std": 2.3,
+            "unique_words": 107,
+            "lexical_diversity": 0.538,
+            "hapax_ratio": 0.332,
+            "avg_word_length": 4.5,
+            "word_length_std": 2.4,
             "transition_rate": 0.0,
             "passive_rate": 0.0,
             "formulaic_rate": 0.0,
-            "punct_ratio": 0.037
+            "punct_ratio": 0.0392
           },
-          "ensemble_score": 0.574,
+          "ensemble_score": 0.5844,
           "ensemble_class": "AI"
         },
         "humanized_agg": {
-          "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,
+          "text_length": 187,
+          "gpt2_perplexity": 14.9,
+          "gpt2_burstiness": 1.0497,
+          "fast_detectgpt": 0.0003,
+          "gptneo_perplexity": 19.2,
+          "char_distribution": 0.6882,
+          "stylometric": 0.2833,
           "stylo_details": {
-            "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,
+            "sentence_count": 9,
+            "sentence_length_mean": 20.8,
+            "sentence_length_std": 11.5,
+            "total_words": 194,
+            "unique_words": 118,
+            "lexical_diversity": 0.608,
+            "hapax_ratio": 0.464,
+            "avg_word_length": 4.1,
+            "word_length_std": 2.2,
             "transition_rate": 0.0,
             "passive_rate": 0.0,
             "formulaic_rate": 0.0,
-            "punct_ratio": 0.0352
+            "punct_ratio": 0.0379
           },
-          "ensemble_score": 0.5844,
-          "ensemble_class": "AI"
+          "ensemble_score": 0.442,
+          "ensemble_class": "Uncertain"
         },
         "humanized_scratch": {
-          "text_length": 119,
-          "gpt2_perplexity": 9.9,
-          "gpt2_burstiness": 0.6588,
-          "fast_detectgpt": 0.001,
-          "gptneo_perplexity": 14.7,
-          "char_distribution": 0.6348,
-          "stylometric": 0.3167,
+          "text_length": 115,
+          "gpt2_perplexity": 13.3,
+          "gpt2_burstiness": 0.5243,
+          "fast_detectgpt": 0.0004,
+          "gptneo_perplexity": 16.6,
+          "char_distribution": 0.6392,
+          "stylometric": 0.3083,
           "stylo_details": {
             "sentence_count": 7,
-            "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,
+            "sentence_length_mean": 16.4,
+            "sentence_length_std": 5.7,
+            "total_words": 117,
+            "unique_words": 84,
+            "lexical_diversity": 0.718,
+            "hapax_ratio": 0.59,
+            "avg_word_length": 4.3,
+            "word_length_std": 2.2,
             "transition_rate": 0.0,
             "passive_rate": 0.0,
             "formulaic_rate": 0.0,
-            "punct_ratio": 0.0416
+            "punct_ratio": 0.0399
           },
-          "ensemble_score": 0.5531,
-          "ensemble_class": "AI"
+          "ensemble_score": 0.4506,
+          "ensemble_class": "Uncertain"
         }
       }
     }
   ],
   "summary": {
-    "ai_mean": 0.473,
-    "humanized_mean": 0.4902,
-    "improvement": -0.0172
+    "ai_mean": 0.4467,
+    "humanized_mean": 0.4765,
+    "improvement": -0.0298
   }
 }