{"dataset": "confit/cremad-parquet", "split": "test", "n": 926, "label_treatment": {"disgust": "excluded", "why": "Disgust is negative but has no dependable arousal direction, so its quadrant is a choice rather than a reading. Excluded here rather than asserting a placement we cannot justify.", "excluded_labels": ["calm", "disgust", "disgusted", "excited", "excitement", "surprise", "surprised"], "n_mapped_labels": 9}, "excluded_labels": ["calm", "disgust", "disgusted", "excited", "excitement", "surprise", "surprised"], "label_mapping": {"anger": "Stressed", "angry": "Stressed", "fear": "Stressed", "fearful": "Stressed", "happy": "Energised", "happiness": "Energised", "sad": "Fatigued", "sadness": "Fatigued", "neutral": "Calm"}, "accuracy": 0.4946, "majority_class_baseline": 0.4125, "beats_baseline": true, "confusion": {"Calm": {"Calm": 80, "Fatigued": 58, "Energised": 15, "Stressed": 10}, "Energised": {"Energised": 108, "Calm": 24, "Stressed": 41, "Fatigued": 17}, "Stressed": {"Stressed": 165, "Fatigued": 62, "Energised": 108, "Calm": 47}, "Fatigued": {"Fatigued": 105, "Calm": 70, "Energised": 11, "Stressed": 5}}, "per_class": {"Calm": {"precision": 0.362, "recall": 0.491, "f1": 0.417, "support": 163}, "Energised": {"precision": 0.446, "recall": 0.568, "f1": 0.5, "support": 190}, "Stressed": {"precision": 0.747, "recall": 0.432, "f1": 0.547, "support": 382}, "Fatigued": {"precision": 0.434, "recall": 0.55, "f1": 0.485, "support": 191}}, "axes": {"arousal_high_vs_low": {"accuracy": 0.7937, "majority_baseline": 0.6177, "lift": 0.176, "n": 926}, "valence_negative_vs_positive": {"accuracy": 0.6091, "majority_baseline": 0.6188, "lift": -0.0097, "n": 926}, "near_chance_axes": ["valence"], "interpretation": "Arousal is recovered well; valence is at or near chance. Since DSI weights valence at 0.45, close to half the index rests on the dimension the model does not resolve - which is the single most important limitation of the affect scale."}, "verdict": "Accuracy 49.5% vs a majority-class baseline of 41.3% - the mapping carries real signal.", "caveat": "CREMA-D is acted, studio-quality American English. This validates the prosody model and our arousal/valence quadrant logic; it does not demonstrate in-domain performance on compressed F1 team radio. The implication runs one way only: a model that cannot recover emotion from clean acted speech is certainly not doing so on radio."}