Mental-Health-Analysis

This model is a fine-tuned version of distilbert-base-uncased on the None dataset. It achieves the following results on the evaluation set:

• Accuracy: 0.9950
• F1: 0.9950
• Loss: 0.0299
• Precision: 0.9950
• Recall: 0.9950

Model description

Mental-Health-Analysis is a transformer-based NLP model fine-tuned from distilbert-base-uncased for mental health text classification.

The model analyzes social media–style text (tweets, Reddit posts, short personal messages) and predicts mental health–related emotional states.
It is designed as a research and educational tool to explore how NLP can support early mental health signal detection.

Supported Classes

• Depression
• Anxiety
• Suicidal Ideation
• Happy
• Neutral / Casual

The project emphasizes ethical AI, social impact, and responsible usage.

Intended uses & limitations

✅ Intended Uses

• Academic research in NLP + mental health
• Educational demonstrations of text classification
• Prototyping mental health–aware AI systems
• Social impact and AI-for-good projects

❌ Out-of-Scope Uses

• Clinical diagnosis
• Medical decision-making
• Crisis intervention without human oversight

⚠️ Disclaimer:
This model is not a substitute for professional mental health care. Predictions are probabilistic signals, not diagnoses.

---

Training and evaluation data

• Data Source: Curated and balanced dataset of social media text
• Platforms: Twitter, Reddit (public posts only)
• Task: 5-class text classification
• Data Split:
  • 80% training
  • 20% evaluation

Basic preprocessing steps:

• Missing text replaced with empty strings
• Text normalized and tokenized using DistilBERT tokenizer
• Padding and truncation applied to max sequence length

---

Training procedure

Model Architecture

• Base Model: distilbert-base-uncased
• Architecture Type: Transformer encoder
• Objective: Multi-class sequence classification (5 labels)

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 5e-05
• train_batch_size: 16
• eval_batch_size: 16
• seed: 42
• optimizer: Use OptimizerNames.ADAMW_TORCH_FUSED with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 500
• num_epochs: 3

Training results

Training Loss	Epoch	Step	Accuracy	F1	Validation Loss	Precision	Recall
0.2938	0.2006	500	0.9574	0.9581	0.1866	0.9612	0.9574
0.119	0.4013	1000	0.9768	0.9769	0.1090	0.9773	0.9768
0.0992	0.6019	1500	0.9819	0.9820	0.0926	0.9822	0.9819
0.0039	0.8026	2000	0.9876	0.9876	0.0731	0.9878	0.9876
0.1583	1.0032	2500	0.9911	0.9911	0.0475	0.9911	0.9911
0.0021	1.2039	3000	0.9915	0.9915	0.0448	0.9915	0.9915
0.033	1.4045	3500	0.9916	0.9916	0.0535	0.9916	0.9916
0.0332	1.6051	4000	0.9915	0.9915	0.0428	0.9915	0.9915
0.0008	1.8058	4500	0.9936	0.9936	0.0360	0.9936	0.9936
0.0263	2.0064	5000	0.9941	0.9941	0.0301	0.9941	0.9941
0.0214	2.2071	5500	0.9939	0.9939	0.0345	0.9939	0.9939
0.0006	2.4077	6000	0.9945	0.9945	0.0302	0.9945	0.9945
0.0114	2.6083	6500	0.9952	0.9952	0.0281	0.9952	0.9952
0.0418	2.8090	7000	0.9950	0.9950	0.0299	0.9950	0.9950

---

How to Use the Model

Quick Start (Pipeline)

from transformers import pipeline

classifier = pipeline(
    "text-classification",
    model="vedabtpatil07/Mental-Health-Analysis"
)

text = "I feel exhausted and hopeless lately."
result = classifier(text)

print(result)

---

Framework versions

• Transformers 4.56.2
• Pytorch 2.8.0+cu126
• Datasets 4.0.0
• Tokenizers 0.22.1

Citation

BibTeX

@misc{mental_health_analysis_2025,
  title        = {Mental Health Analysis: DistilBERT-Based Text Classification},
  author       = {Vedant Patil and Ansh Jaiswal},
  year         = {2025},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/vedabtpatil07/Mental-Health-Analysis}}
}

Model Card Authors

• Vedant Patil
• Ansh Jaiswal