metadata

library_name: transformers
tags:
  - MentalHealth
  - Counseling
  - Chatbot
  - InstructionTuning
  - QLoRA
datasets:
  - Amod/mental_health_counseling_conversations
language:
  - en
base_model:
  - TinyLlama/TinyLlama-1.1B-Chat
pipeline_tag: text-generation

Model Card: Mental Health Counselor Chatbot (TinyLlama-1.1B)

This model is a lightweight mental health chatbot built on TinyLlama-1.1B-Chat, fine-tuned using QLoRA on the Amod/mental_health_counseling_conversations dataset.

⚠️ Note: This model was fine-tuned on Google Colab (free T4 GPU) for only 1 epoch, intended as a test to evaluate the ability of TinyLlama to respond to counseling prompts.
🧠 Performance can significantly improve with longer training, more data, and better hyperparameter tuning.

Model Details

Model Description

Model Type: Causal Language Model (Instruction-Tuned)
Base Model: TinyLlama/TinyLlama-1.1B-Chat
Fine-tuned by: Ali Haider
Dataset: Amod/mental_health_counseling_conversations
Language: English
License: Apache 2.0

⚠️ This is a prototype model. It was fine-tuned using only 1 epoch on a small sample dataset for demonstration and testing purposes.

Uses

Direct Use

For generating supportive and empathetic responses to mental health-related user inputs. Useful for:

Mental health Q&A bots
Conversational agents in wellness apps

Out-of-Scope Use

Not a substitute for licensed therapy.
Should not be used for clinical decisions or crisis support.

Bias, Risks & Limitations

The model may produce biased or generic responses.
Only trained on one small dataset, so coverage is limited.
May hallucinate or offer vague advice if prompted outside of its domain.

How to Use

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("ali1001/mental-health-tinyllama-bot")
tokenizer = AutoTokenizer.from_pretrained("ali1001/mental-health-tinyllama-bot")

prompt = "I'm feeling very anxious lately and can't sleep. What should I do?"
inputs = tokenizer(prompt, return_tensors="pt")
output = model.generate(**inputs, max_new_tokens=150)
print(tokenizer.decode(output[0], skip_special_tokens=True))