Mistral-7B-Instruct-v0.3-Mental-Health-chatbot

This is a fine-tuned version of Mistral-7B-Instruct-v0.3 for empathetic and mental-health-oriented dialogue.
It has been adapted to provide supportive, safe, and context-aware responses in conversations related to mental health.

Disclaimer:
This model is not a replacement for professional mental health care. It should only be used for research and educational purposes.
If you are experiencing a crisis, please seek immediate help from a qualified professional.

---

Model Details

• Base model: Mistral-7B-Instruct-v0.3
• Fine-tuned for: Empathetic dialogue, mental health supportive responses
• Framework: Transformers

---

Datasets Used

This model was fine-tuned on multiple open-source mental health conversation datasets:

1. Amod/mental_health_counseling_conversations

   • Real counseling Q&A with licensed professionals
   • JSON format, ~1K–10K entries

2. mpingale/mental-health-chat-dataset

   • Tabular chat dataset with question/answer pairs and therapist info
   • Parquet format, ~2.78K entries

3. heliosbrahma/mental_health_chatbot_dataset

   • Small conversational dataset (~172–1K entries)
   • Q&A pairs for chatbot training, MIT license

4. marmikpandya/mental-health

   • Large dataset of ~13.4K entries
   • JSON format, mental health dialogues

By combining these datasets, the model learns empathetic, contextually relevant, and supportive responses.

---

Intended Use

• Research in empathetic AI
• Exploring AI for mental health support
• Developing chatbots that respond with empathy

Not intended for:

• Real-time crisis management
• Replacing human therapists or counselors

---

Evaluation Results

The model was evaluated using multiple NLP metrics to measure fluency, relevance, and semantic similarity:

📊 Automatic Metrics

• BERTScore (F1 mean): 0.876

• SBERT Cosine Similarity (mean): 0.695

• ROUGE Scores:

  • ROUGE-1: 0.341
  • ROUGE-2: 0.135
  • ROUGE-L: 0.221
  • ROUGE-Lsum: 0.225

• BLEU: 0.081

• chrF++: 38.05

📉 Perplexity

• Base Model Perplexity: 11.32
• Fine-tuned Model Perplexity: 2.52

Lower perplexity indicates the fine-tuned model produces more coherent and predictable text compared to the base model.

These results show that fine-tuning substantially improved the model’s language generation quality, making it better aligned for empathetic and supportive mental health conversations.

---

Example Usage

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

# Load your model and tokenizer
model_name = "Tanneru/Mistral-7B-Instruct-v0.3-Mental-Health-chatbot"
model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype=torch.float16, device_map="auto")
tokenizer = AutoTokenizer.from_pretrained(model_name)

# Define the chat
chat = [
    {"role": "system", "content": "You are EmpathAI, a supportive and empathetic AI trained for mental health conversations. Always respond with kindness and empathy.Do not use hashtags or excessive emojis.Do not repeat phrases or sentences"},
    {"role": "user", "content": "I feel very anxious about my exams. Can you help me calm down?"}
]

# Tokenize input
inputs = tokenizer.apply_chat_template(
    chat,
    add_generation_prompt=True,
    return_tensors="pt"
).to(model.device)

# Generate response
outputs = model.generate(
    inputs,
    max_new_tokens=256,
    do_sample=True,
    temperature=0.6,
    top_p=0.9,
    no_repeat_ngram_size=6,     # hard block repeated 6-gram
    repetition_penalty=1.15,    # soft discouragement of reuse
    pad_token_id=tokenizer.eos_token_id
)

# Extract assistant's reply
response = outputs[0][inputs.shape[-1]:]
print("EmpathAI:", tokenizer.decode(response, skip_special_tokens=True))

---

Limitations

• The model may sometimes generate inaccurate or harmful advice.
• Responses may vary depending on phrasing and context.
• Should not be solely relied upon for medical or therapeutic guidance.

---

Citation

If you use this model in your research or project, please cite it:

@misc{tanneru2025mistralmentalhealth,
  title   = {Mistral-7B-Instruct-v0.3-Mental-Health-chatbot},
  author  = {Tanneru},
  year    = {2025},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/Tanneru/Mistral-7B-Instruct-v0.3-Mental-Health-chatbot}}
}