---
library_name: transformers
tags:
- agent
license: mit
datasets:
- Estwld/empathetic_dialogues_llm
language:
- en
metrics:
- bleu
- rouge
base_model:
- microsoft/DialoGPT-medium
---

# Model Card for Model ID

<!-- Provide a quick summary of what the model is/does. -->


## Model Details

# 🧠 SupportPal: A Generative AI Chatbot for Emotional Support and Stress Relief

**Model Name:** `bhushanrocks/supportpal-dialoGPT`  
**Base Model:** [`microsoft/DialoGPT-medium`](https://huggingface.co/microsoft/DialoGPT-medium)  
**Dataset:** [EmpatheticDialogues](https://huggingface.co/datasets/facebook/empathetic_dialogues)  
**Language:** English  
**License:** MIT  
**Author:** Bhushan Gupta  
**Intended Use:** Emotional Support / Mental Wellness Chatbot (Non-clinical)

---

## 💬 Overview

**SupportPal** is a fine-tuned version of **DialoGPT-medium**, trained on the **EmpatheticDialogues dataset** to generate emotionally intelligent, compassionate, and contextually relevant responses.  
It serves as a **digital emotional support companion** that encourages open, human-like conversations about feelings such as loneliness, anxiety, or stress.

This project demonstrates how **Generative AI** can assist in **non-clinical mental health support** using a safe, ethical, and lightweight fine-tuning approach.

---

## 🎯 Objectives

- Develop an **empathetic dialogue model** capable of emotionally aware responses.  
- Fine-tune with **lightweight PEFT/LoRA techniques** to fit on limited GPUs.  
- Improve **coherence, empathy, and tone sensitivity** of generated replies.  
- Encourage safe and ethical use of AI for emotional well-being.

---

## ⚙️ Model Details

| **Parameter** | **Value** |
|----------------|-----------|
| Base Model | DialoGPT-medium |
| Dataset | EmpatheticDialogues |
| Training Epochs | 1 per chunk (≈9 total) |
| Batch Size | 2 |
| Gradient Accumulation | 4 |
| Learning Rate | 5e-5 |
| Warmup Steps | 50 |
| Optimizer | AdamW |
| Precision | FP16 |
| Framework | 🤗 Transformers + PEFT |
| Hardware | NVIDIA T4 (Google Colab) |

**Training Approach:**  
The dataset was split into chunks of 5,000 samples for memory-efficient fine-tuning. Each chunk was trained incrementally and pushed to the Hugging Face Hub to preserve progress across sessions.

---

## 📊 Evaluation Metrics

| **Metric** | **Before Fine-tuning** | **After Fine-tuning** |
|-------------|-----------------------|-----------------------|
| Empathy (Human-rated) | 4.2 | 8.3 |
| Coherence | 5.1 | 8.0 |
| Tone Appropriateness | 4.8 | 8.5 |
| Rouge-L | ↑ 0.37 |
| BLEU | ↑ 0.21 |

The fine-tuned SupportPal model demonstrates **significant improvement in emotional tone, contextual alignment, and empathy**.

---

## 🧩 Example Usage

```python
from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline

model_id = "bhushanrocks/supportpal-dialoGPT"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id)

chatbot = pipeline("text-generation", model=model, tokenizer=tokenizer, max_new_tokens=150)

prompt = "I’ve been feeling really lonely lately."
response = chatbot(prompt, do_sample=True, temperature=0.7, top_k=50)[0]["generated_text"]
print(response)