spambert-tinybert-real-plus-synth-v0.9-thr057

TinyBERT spam classifier trained from prancyFox/tiny-bert-enron-spam with mixed real and synthetic spam data.

Training recipe

• Base model: prancyFox/tiny-bert-enron-spam
• Real data: train/train_processed.csv
• Synthetic augmentation: synthetic-to-real ratio 0.9
• Selected decision threshold: 0.571892

Full real-data evaluation

Evaluation file: eval_full_real_threshold_0p571892.json

• Accuracy: 0.9769
• Macro F1: 0.9305
• Spam precision: 0.9574
• Spam recall: 0.8036
• Spam F1: 0.8738
• Confusion matrix [ham, spam]: [[504, 2], [11, 45]]

Validation threshold tuning

Threshold tuning file: threshold_tuning.json

• Recommended threshold: 0.571892
• Validation spam precision: 1.0000
• Validation spam recall: 0.7000
• Validation spam F1: 0.8235