qwen2.5-1.5b-ticket-classifier

Model Card for Model ID

🚀 Production-Grade SLM Structured Data Extractor

Fine-Tuned Qwen-1.5B-Instruct for Zero-Yapping Strict JSON Schema Compliance

---

📌 Business Case & Problem Statement

In enterprise production environments, unstructured text (customer support tickets, invoices, logs) must be mapped to structured databases (SQL/NoSQL) with 100% deterministic reliability.

While massive commercial LLMs (like GPT-4o) achieve high accuracy, they introduce major bottlenecks for high-throughput narrow tasks:

1. High API Costs: Processing millions of tokens daily is economically unsustainable.
2. High Latency: Cloud API round-trips slow down real-time automated routing workflows.
3. Data Privacy Risks: Transmitting sensitive client information (PII) to third-party APIs violates data compliance laws (GDPR/HIPAA).
4. Formatting Violations ("Yapping"): General-purpose base models frequently violate strict formatting constraints by adding conversational filler (e.g., "Here is the JSON you requested...").

The Solution

This project demonstrates Cost-Conscious AI Engineering by fine-tuning a 1.5 Billion parameter Small Language Model (SLM) to perform production-grade, schema-bound classification with near-zero latency and fraction of a cent compute cost, making it entirely deployable on the edge or low-cost commodity GPUs.

---

📊 Performance & ROI Benchmark

We evaluated the model on a hidden test dataset of customer support tickets across 4 categorical schema keys: category, urgency, sentiment, and action_required.

💰 Estimated Cost Breakdown (At Scale)

Assumed volume: 1 Million Tickets/Month (~300M Tokens processed)

• Commercial LLM API (e.g., GPT-4o Class): ~$750 - $1,500 / month
• This Fine-Tuned SLM (Hosted on Single Low-End Instance / Serverless GPU): <$25 / month
• Net Business Savings: ~96.5% Cost Reduction with 100% data privacy.

---

🛠️ Technical Implementation & Architecture

1. Technology Stack

• Base Model: Qwen/Qwen2.5-1.5B-Instruct
• Fine-Tuning Technique: Parameter-Efficient Fine-Tuning (PEFT) using QLoRA (4-bit quantization)
• Infrastructure: Trained via PyTorch and Hugging Face Trainer pipeline on a single consumer-grade T4 GPU.
• Inference UI: Gradio application deployed seamlessly on Hugging Face Spaces.

2. Target Features & Schema Constraints

The model was fine-tuned using custom-engineered synthetic dataset structures following the ChatML format to map arbitrary inputs directly into this immutable JSON schema:

{
  "category": "billing | technical | refund | account | general",
  "urgency": "low | medium | high",
  "sentiment": "positive | negative | neutral",
  "action_required": "auto_reply | escalate | close"
}