mdk615661/it-helpdesk-qlora-v4

IT Helpdesk QLoRA Adapter — v4

LoRA adapter for IT helpdesk ticket classification. Load this adapter on top of mdk615661/it-helpdesk-merged-v3.

Model Details

• Type: QLoRA Adapter (PEFT)
• Base Model: mdk615661/it-helpdesk-merged-v3
• LoRA: r=16, alpha=32
• Training Data: 2,000 IT helpdesk records
• Training Loss: 0.187 (3 epochs)
• Hardware: Kaggle T4 GPU (33 min)

Usage

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

base_model = AutoModelForCausalLM.from_pretrained(
    "mdk615661/it-helpdesk-merged-v3",
    dtype=torch.float16,
    device_map="auto"
)
model = PeftModel.from_pretrained(base_model, "mdk615661/it-helpdesk-qlora-v4")
tokenizer = AutoTokenizer.from_pretrained("mdk615661/it-helpdesk-merged-v3")

prompt = """### Instruction:
Normalize and classify this IT helpdesk ticket.

### Input:
Laptop is not turning on

### Output:
"""

inputs  = tokenizer(prompt, return_tensors="pt").to("cuda")
outputs = model.generate(**inputs, max_new_tokens=150, do_sample=False)
print(tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True))

Output Format

Category: Hardware SubCategory: Hardware - Laptop Normalized: laptop not working Priority: Medium Insight: Hardware failure preventing user from working. Recommendation: Raise repair request with IT hardware team.

Categories

Category	Subcategories
Hardware	Laptop, Charger, Mobile
Software	Installation, VPN, Password Reset, O365, Teams, MFA Reset
Incident	Critical, Network Outage, Security, Service Outage, Performance
Procurement	Hardware, Software
Onboarding & Offboarding	Onboarding, Offboarding
Cloud & Infrastructure	DR/BCP, Network Config, System Config
Asset	Asset Management, Asset Request
Others	Account Management, Audit, Change Management

Version History

Version	Data	Loss
v3	1,141 real TruMIS tickets	—
v4 (this)	+ 2,000 Qwen records	0.187

Training Hyperparameters

• Epochs: 3
• Batch size: 4
• Gradient accumulation: 4
• Learning rate: 2e-4
• Warmup steps: 50
• LR scheduler: cosine
• Precision: fp16