EndConvo-health-deberta-v2

EndConvo-health-deberta-v2

Model Description

The EndConvo-health-deberta-v2 is a fine-tuned conversational AI model based on the DeBERTa architecture. It is designed for binary classification tasks to determine whether a conversation in a health-related chatbot has reached its endpoint or should continue. The model significantly improves efficiency by identifying conversation closure, especially in healthcare applications, where accurate and timely responses are crucial.

---

Intended Use

• Primary Use Case: End-of-conversation detection in health-related chatbot systems.
• Scope of Application: Healthcare dialogues, customer support automation, or any domain requiring conversational flow control.
• Limitations:
  • Reduced recall for the "True" (conversation ending) class, which could affect performance in ambiguous scenarios.
  • The model requires GPU support for efficient inference on large-scale data.

---

Training

• Structure: Binary classification dataset with labels:
  • 0 for "Continue conversation"
  • 1 for "End conversation."
• Size: 4,000 training samples and 1,000 validation samples.
• Source: Annotated conversational data designed for healthcare-related use cases.
• Preprocessing:
  • Tokenization using DeBERTa tokenizer.
  • Maximum sequence length of 256 tokens.
  • Truncation applied for longer conversations.

---

Model Details

• Base Model: DeBERTa-V2
• Training Framework: Hugging Face Transformers
• Optimizer: AdamW with weight decay
• Loss Function: Cross-entropy loss
• Batch Size: 16
• Epochs: 3
• Learning Rate: 5e-5
• Evaluation Metric: Accuracy, Precision, Recall, F1-score

---

Evaluation Metrics

• Overall Accuracy: 86.6%
• Precision: 86.7%
• Recall: 58.0%
• F1-Score: 69.5%
• Validation Loss: 0.3729

Confusion Matrix

• True Negatives (TN): 71.29%
• False Positives (FP): 2.35%
• False Negatives (FN): 11.06%
• True Positives (TP): 15.29%

Detailed Report

---

Pipeline and Usage

• Task Type: Text classification for conversation flow.
• Pipeline: Predicts whether a conversation should continue or end.

Example Usage

from transformers import AutoTokenizer, AutoModelForSequenceClassification

# Load tokenizer and model
tokenizer = AutoTokenizer.from_pretrained("MathewManoj/EndConvo-health-deberta-v2")
model = AutoModelForSequenceClassification.from_pretrained("MathewManoj/EndConvo-health-deberta-v2")

# Example text input
text = "Thank you for your help. I don't have any more questions."

# Tokenize the input
inputs = tokenizer(text, return_tensors="pt")
outputs = model(**inputs)

# Prediction
prediction = outputs.logits.argmax(dim=-1).item()
print("Prediction:", "End" if prediction == 1 else "Continue")

---

Performance Insights

Strengths:

• High accuracy and precision indicate the model performs well in correctly identifying most "Continue" conversations.

Limitations:

• Lower recall for "End" conversations suggests the need for additional data augmentation or fine-tuning to improve sensitivity.

---

Environment and Dependencies

• Framework: Hugging Face Transformers (v4.46.3)
• Python Version: 3.8+
• Dependencies:
  • torch
  • transformers
  • safetensors
  • numpy

Conda Environment Configuration

name: huggingface-env
channels:
  - defaults
  - conda-forge
dependencies:
  - python=3.8
  - pip
  - pip:
      - torch==2.4.1
      - transformers==4.46.3
      - safetensors

---

Model Limitations

1. The model exhibits reduced recall for the "End conversation" class, which could impact its utility in edge cases.
2. Requires labeled data for fine-tuning in other domains or applications.