base_model: "answerdotai/ModernBERT-base"
HAV classifier (modernbert-base)
Binary: High Analytical Value (HAV) vs Low (LAV) for social posts. Precision-optimised (F0.5).
Data preparation
Model has been trained on data with spam removed, spam posts could be classified as HAV. Not trained on tiktok data.
Operating point
Apply a decision threshold of 0.909 on softmax(logits)[:, 1].
Do NOT use argmax@0.5 — this model is precision-tuned and 0.5 gives far lower precision.
Metrics
"objective": {
"metric": "F0.5 (HAV)",
"beta": 0.5,
"min_hav_recall": 0.3,
"decision_threshold": 0.9099,
"val_selection": {
"threshold": 0.9099,
"hav_precision": 0.8238,
"hav_recall": 0.5035,
"nonhav_recall": 0.9753,
"fbeta": 0.7308,
"beta": 0.5
}
},
"data": {
"train_rows": 10866,
"train_hav": 2033,
"eval_rows": 2366,
"eval_hav": 440
},
"test_overall": {
"accuracy": 0.8762,
"hav_f05": 0.6841,
"hav_f1": 0.5936,
"macro_f1": 0.7603,
"weighted_f1": 0.865,
"average_precision": 0.7163
},
"test_per_class": {
"non_HAV": {
"precision": 0.8916,
"recall": 0.9652,
"f1-score": 0.927,
"support": 1926.0
},
"HAV": {
"precision": 0.7616,
"recall": 0.4864,
"f1-score": 0.5936,
"support": 440.0
}
}
Caveats
Misclassfied negative class still an issue. pos_weight=4.9, max_length=256.
