Fine-tuned version of mental-bert-base-uncased for 4-class mental health severity classification. Built as part of the EchoCare project (M1 — NLP Core).

Task

Given a user's text input, classify it into one of 4 severity levels:

Label	Class	Description
0	Low	Mild stress, general low mood
1	Medium	Moderate depression symptoms
2	High	Severe depression, high distress
3	Crisis	Active suicidal ideation, immediate risk

Usage

python

from transformers import BertTokenizerFast, BertForSequenceClassification
import torch

tokenizer = BertTokenizerFast.from_pretrained("Aalia-Laghari/hamdardai-bert")
model     = BertForSequenceClassification.from_pretrained("Aalia-Laghari/hamdardai-bert")
model.eval()

text   = "I haven't left my room in days and nothing feels real anymore"
inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=128)

with torch.no_grad():
    logits = model(**inputs).logits
    pred   = logits.argmax(dim=-1).item()

labels = ["Low", "Medium", "High", "Crisis"]
print(labels[pred])

Training Data

Three datasets combined and preprocessed:

Dataset	Samples	Classes used
Depression Severity (Reddit)	3,549	Low, Medium, High, Crisis
SDCNL	1,820	Medium, Crisis
SuicideWatch	5,000 (sampled)	Low, Crisis
Total	~11,150 (after augmentation)

Class distribution after augmentation:

Low: 5,113 | Medium: 2,000 | High: 1,200 | Crisis: 3,638

Training Details

Parameter	Value
Base model	mental/mental-bert-base-uncased
Max sequence length	128
Batch size	32
Learning rate	2e-5
LR scheduler	Cosine with warmup (15%)
Weight decay	0.05
Label smoothing	0.1
Loss function	Ordinal CrossEntropy + MSE (λ=0.5)
Dropout	0.3 (all layers)
Early stopping	Patience 5 on val macro-F1
Max epochs	20
Data augmentation	Synonym swap + pattern paraphrase on minority classes

Performance (Test Set)

Metric	Score
Accuracy	81.6%
Macro F1	0.764
Precision	0.755
Recall	0.775

Per-class breakdown:

Class	Precision	Recall	F1	Support
Low	0.884	0.865	0.875	512
Medium	0.508	0.508	0.508	120
High	0.796	0.908	0.848	120
Crisis	0.832	0.819	0.825	364

Limitations

Medium class F1 (0.508) is lower due to inherent ambiguity between mild and moderate severity
Trained on English Reddit text — may not generalise to formal clinical language
Not a substitute for professional mental health assessment
Inter-annotator agreement on these datasets is ~70–80%, setting a natural ceiling on accuracy

Intended Use

Research and educational purposes as part of the EchoCare mental health support application. Not intended for clinical diagnosis or standalone mental health intervention.

Authors

Aalia Laghari — NLP Core (M1), EchoCare Project