Billing SLM - Fine-Tuning Llama 3.2 3B Instruct Model on Telecom Billing Data

Overview

This project utilizes customer billing data to fine-tune a Llama 3B Instruct model, with the goal of enhancing the model's ability to understand and generate responses relevant to billing inquiries. The dataset comprises detailed billing records, including charge breakdowns, event tracking, installment payments, and promotional credits. The data will allow the model to handle typical customer questions about billing, providing clear, accurate, and personalized responses.

Project Goals

Enhance Customer Service: Train the model to understand customer billing inquiries and generate accurate, context-aware responses. Improve Data Interpretation: Enable the model to extract specific details (e.g., plan charges, device payments, and surcharges) and respond to nuanced questions about billing discrepancies, promotional credits, and installment plans. Support Billing Insights: Train the model to answer common billing questions, such as breaking down monthly charges, understanding charge adjustments, and identifying recurring versus one-time charges.

Data Structure

The dataset consists of JSON files representing individual customer billing synthetic records over multiple billing cycles. Each record includes:

  • Customer ID: A unique identifier for each customer.
  • Billing Date: The date of each billing cycle.
  • Previous and Current Bill Amounts: Tracking bill amounts over time.
  • Account Level Charges: Any charges applied to the overall account.
  • Mobile Number Charges: Detailed breakdown per mobile number, including:
  • Events: Notable billing events (e.g., "Device Not Returned").
  • Sections: Categorized charges such as one-time fees, plans, devices, perks, surcharges, and taxes.
  • Itemized Device Details: Information on installment payments, promotional credits, and remaining balance.

Fine-Tuning Process

  • Data Formatting: Convert the JSON data into a suitable format for Llama 3B fine-tuning, where each entry maps customer questions to expected answers, covering all possible scenarios within the billing data.
  • Instruction Formatting: Craft prompt-response pairs based on real-life billing inquiries to help the model learn to identify and answer specific questions.
  • Model Training: Fine-tune the Llama 3B model with a focus on customer service scenarios, ensuring it can handle complex billing inquiries accurately and respond with the appropriate details.
  • Evaluation & Testing: Validate model responses with additional billing questions to assess accuracy, clarity, and relevance. Iterate on training data to refine responses as needed.

Usage: The model can be integrated into customer support chatbots or virtual assistants to assist with billing inquiries, helping customers get quick, accurate answers to questions about their bill breakdowns, promotional credits, installment plans, and other related topics.

Example Prompts

  • "Can you explain why my current bill is higher than last month?"
  • "What are the details of the device charges on my bill last month?"
  • "How many payments remain for my device installment?"
  • "Why didn’t I receive my trade-in promotion this month?"

By leveraging this fine-tuned model, customer service teams can provide a more efficient, automated solution to handle a wide range of billing inquiries, enhancing the overall customer experience.

Downloads last month
33
Safetensors
Model size
3.21B params
Tensor type
BF16
·
Inference API
Unable to determine this model's library. Check the docs .

Model tree for subhashtalluri/billing-slm-finetuned-3-2-llama3b

Finetuned
(166)
this model