Pulastya0/Deberta-V5-BEST-Employee-Final

🤖 DeBERTa-v3-base for Employee IT Support Ticket Classification

📖 Model Overview

This model is a fine-tuned version of microsoft/deberta-v3-base for classifying employee IT support tickets into 11 categories.
It was trained in two stages:

1. Domain Adaptation — fine-tuned on ~12k general customer support tickets.
2. Task Adaptation — fine-tuned on 2.5k synthetic employee IT tickets.

The model automates helpdesk ticket routing by predicting the correct support category.

---

🗂️ Labels

The model predicts one of the following categories:

• Network
• Software
• Account
• Training
• Security
• Licensing
• Communication
• RemoteWork
• Hardware
• Infrastructure
• Performance

---

🎯 Intended Uses

• Automated Ticket Routing — Assign new tickets to the right IT team.
• Helpdesk Analytics — Analyze ticket trends.
• Chatbots — Suggest relevant answers or knowledge base articles.

⚠️ Limitations:

• Synthetic training data may not capture all company-specific jargon.
• Validation accuracy is near-perfect, but real-world accuracy expected is 85–95%.

---

💻 Usage

from transformers import pipeline

# Load model
classifier = pipeline("text-classification", model="your-username/deberta-it-support")

subject = "VPN connection dropping"
description = "My VPN disconnects every 15 minutes, preventing access to remote servers."

text_input = f"[SUBJECT] {subject} [TEXT] {description}"

result = classifier(text_input)
print(result)
# [{'label': 'RemoteWork', 'score': 0.98}]

🏋️Training Data

Stage	Dataset	Size	Purpose
Stage 1	Customer Support Tickets (public)	~12,000	Domain Adaptation
Stage 2	Synthetic Employee IT Tickets	2,500	Task Adaptation

Hyperparameters

Hyperparameter	Stage 1	Stage 2
Learning Rate	2e-5	5e-6
Epochs	3	5
Batch Size (per device)	8	8
Gradient Accumulation	4	4
Optimizer	AdamW	AdamW
Precision	FP16	FP16

📊 Evaluation

The final model achieved 99.4% accuracy on the validation split of the synthetic dataset. The best checkpoint was saved using the load_best_model_at_end strategy, based on validation loss. As noted in the limitations, real-world performance will likely be slightly lower but is expected to be high.

This model was fine-tuned by [Pulastya/Pulastya0].

Base model microsoft/deberta-v3-base is provided under the MIT license.