Table 3.

Summary of Seven Mental Health Datasets Employed for Our Experiment. The top four datasets are used for both training and testing, while the bottom three datasets are used for external evaluation. We define six diverse mental health prediction tasks on these datasets.

Dataset	Task	Dataset Size	Text Length (Token)
Dreaddit [120] Source: Reddit	#1: Binary Stress Prediction post-level	Train: 2838 (47.6% False, 52.4% True) Test: 715 (48.4% False, 51.6% True)	Train: 114 ± 41 Test: 113 ± 39
DepSeverity [80] Source: Reddit	#2: Binary Depression Prediction post-level	Train: 2842 (72.9% False, 17.1% True) Test: 711 (72.3% False, 17.7% True)	Train: 114 ± 41 Test: 113 ± 37
	#3: Four-level Depression Prediction post-level	Train: 2842 (72.9% Minimum, 8.4% Mild, 11.2% Moderate, 7.4% Severe) Test: 711 (72.3% Minimum, 7.2% Mild, 11.5% Moderate, 10.0% Severe)	Train: 114 ± 41 Test: 113 ± 37
SDCNL [49] Source: Reddit	#4: Binary Suicide Ideation Prediction post-level	Train: 1516 (48.1% False, 51.9% True) Test: 379 (49.1% False, 50.9% True)	Train: 101 ± 161 Test: 92 ± 119
CSSRS-Suicide [40] Source: Reddit	#5: Binary Suicide Risk Prediction user-level	Train: 400 (20.8% False, 79.2% True) Test: 100 (25.0% False, 75.0% True)	Train: 1751 ± 2108 Test: 1909 ± 2463
	#6: Five-level Suicide Risk Prediction user-level	Train: 400 (20.8% Supportive, 20.8% Indicator, 34.0% Ideation, 14.8% Behavior, 9.8% Attempt) Test: 100 (25.0% Supportive, 16.0% Indicator, 35.0% Ideation, 18.0% Behavior, 6.0% Attempt)	Train: 1751 ± 2108 Test: 1909 ± 2463
Red-Sam [105] Source: Reddit	#2: Binary Depression Prediction post-level	External Evaluation: 3245 (26.1% False, 73.9% True)	External Evaluation: 151 ± 139
Twt-60Users [56] Source: Twitter	#2: Binary Depression Prediction post-level	External Evaluation: 8135 (90.7% False, 9.3% True)	External Evaluation: 15 ± 7
SAD [76] Source: SMS-like	#1: Binary Stress Prediction post-level	External Evaluation: 6185 (6.0% False, 94.0% True)	External Evaluation: 13 ± 6