Model Card for qwen2.5-mentalchat16k

This model is a LoRA fine-tuned version of Qwen/Qwen2.5-1.5B-Instruct trained on the ShenLab/MentalChat16K dataset for mental health conversations. It introduces low-rank adapters to query and value projections while keeping the base model frozen, allowing task-specific adaptation with minimal compute overhead.

Model Details

Model Description

This model is designed to generate context-aware, mental-health-sensitive responses in English. It fine-tunes a large pre-trained language model (Qwen2.5-1.5B-Instruct) using LoRA adapters to focus on mental health conversational patterns. The adapters are lightweight, targeting the attention mechanism, and maintain most of the base model frozen.

• Developed by: Rohan Italiya, Janvi Patel

• Model type: Transformer-based Causal Language Model

• Language(s): English

• Finetuned from model : Qwen/Qwen2.5-1.5B-Instruct

Model Sources [optional]

• Repository: https://huggingface.co/rohi1810/qwen2.5-mentalchat16k

Uses

Direct Use

• Generate mental-health-aware conversational responses.
• Useful for research, educational purposes, and building chatbots with sensitive conversational context.
• Can be used directly via Python using the Transformers library.

Downstream Use [optional]

• Can be further fine-tuned for specialized subdomains or additional mental health datasets.
• Suitable for integration into larger applications that require contextual mental health responses.

Out-of-Scope Use

• Not suitable for medical advice, diagnosis, or therapy.
• Outputs may contain biases, inaccuracies, or inappropriate suggestions outside the dataset domain.
• Should not be used in high-stakes decision-making without human oversight.

Bias, Risks, and Limitations

• Trained on MentalChat16K, which may reflect biases present in the dataset.
• Responses are generated and may not always be safe, accurate, or appropriate.
• Limited to English conversational context.
• LoRA adapters reduce overfitting but may underperform on completely unseen scenarios.

Recommendations

• Validate outputs before deploying in any sensitive application.
• Implement filtering or human oversight for high-risk situations.
• Users should be aware of dataset limitations and model biases.

How to Get Started with the Model

from transformers import AutoTokenizer, AutoModelForCausalLM, pipeline
import torch

# Path to your fine-tuned model
finetuned_path = "rohi1810/qwen2.5-mentalchat16k"

# Load tokenizer and model
tokenizer_main = AutoTokenizer.from_pretrained(finetuned_path)
model_main = AutoModelForCausalLM.from_pretrained(
    finetuned_path,
    dtype=torch.float16,  # reduce memory usage
    device_map="auto"     # automatically place layers on available GPU(s)
)

# Create a text generation pipeline
generator = pipeline("text-generation", model=model_main, tokenizer=tokenizer_main)

# Example prompt
prompt = "<s>[INST] User: I've been struggling with my identity as a member of the LGBTQ community. It's been really hard for me to come to terms with who I am and how society perceives me. I feel like I'm constantly judged and misunderstood, which has taken a toll on my mental health. I need help navigating through these challenges and finding acceptance within myself. [/INST]"

# Generate response
output = generator(prompt, max_new_tokens=150)

# Extract generated text
response = output[0]["generated_text"].split("[/INST]")[-1].strip()
print(response)