llama2-SuperHornet

Model Card for IT Helpdesk Ticket Support Llama-2 Model

🧠 Model Summary

This model is a fine-tuned version of Meta’s Llama-2-7B-Chat trained on a synthetic IT Helpdesk Ticket dataset containing 1500 real-like corporate support tickets.
The model is designed to understand and answer IT support–related questions in natural language — such as password issues, VPN problems, hardware failures, or system access requests.

It can be used for:

• Automated helpdesk assistants
• FAQ retrieval and response systems
• Internal IT ticket triaging tools
• Smart support chatbots

---

📄 Model Details

• Base model: meta-llama/Llama-2-7b-chat-hf
• Fine-tuned by: Dharunpandi
• Language(s): English
• Model type: Causal language model (decoder-only Transformer)
• License: Meta Llama 2 Community License
• Finetuned on: Synthetic IT Helpdesk Ticket dataset
• Task: Text generation / Question answering in helpdesk context
• Framework: 🤗 Transformers

---

💡 Intended Use

✅ Direct Use

• Generate concise and accurate answers to IT support–related queries.
• Suggest relevant departments or assigned teams for reported issues.
• Summarize or classify IT support ticket content.

🔧 Downstream Use

• Integrate with IT Service Management (ITSM) tools like Jira, Freshservice, or ServiceNow.
• Deploy as a chatbot or assistant inside company portals.

🚫 Out-of-Scope Use

• Non-technical or non-English queries.
• High-stakes decision-making without human oversight.
• Retrieval of private or sensitive data.

---

⚠️ Bias, Risks, and Limitations

• The model was trained on synthetic data, not real corporate tickets — accuracy may vary in real environments.
• Some responses might include hallucinated metadata (e.g., employee names or departments).
• Should not be used for security-sensitive IT operations (e.g., password resets without validation).

---

⚙️ Training Details

Dataset

• Name: Synthetic IT Helpdesk Ticket Dataset
• Format: JSON (fields include Ticket_ID, Title, Description, Department, Assigned_Team, Created_At, Updated_At)
• Size: 1500 samples

Preprocessing

• Combined Title and Description into unified text input.
• Removed random identifiers like Ticket_ID and timestamps to reduce noise.

Training Hyperparameters

• Epochs: 3
• Learning rate: 1e-4
• Batch size: 2
• Optimizer: paged_adamw_8bit
• Precision: fp16 mixed
• Max steps: 20
• Save strategy: epoch

---

🧩 Evaluation

Metrics

• Qualitative evaluation via generated responses for unseen IT queries.
• Example:
  • User: “How do I reset my corporate email password?”
  • Model Response: “You can reset your password through the corporate login portal under ‘Forgot Password?’ or contact the IT Helpdesk team.”

Observations

• The model can generalize to unseen support scenarios.
• May occasionally add extra fields like Department or Employee Name.

---

🌱 Environmental Impact (Approximation)

• Hardware: NVIDIA A100 (40GB)
• Training Time: ~2 GPU hours
• Estimated CO₂ emissions: < 1 kg CO₂eq
• Compute provider: Google Colab

---

🔍 How to Use the Model

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_name = "your-username/it-helpdesk-llama2"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name).to("cuda")

question = "How do I reset my corporate email password?"
prompt = f"Question: {question}\nAnswer concisely and helpfully:\n"

inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
outputs = model.generate(**inputs, max_new_tokens=200)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))