PayFlow Hallucination-Controlled Billing Assistant

Table of Contents

1. Project Overview
2. Problem Statement
3. Why Hallucinations Are Dangerous in Billing Systems
4. Project Goal
5. Model & Training Strategy
6. Dataset Design
7. Policy-Driven Training Approach
8. Training Process (Step-by-Step)
9. Evaluation Methodology
10. Results Summary
11. Challenges Faced
12. Key Learnings
13. Limitations & Future Improvements
14. Training Configuration Notes
15. Conclusion

---

1. Project Overview

This project focuses on reducing hallucinations in a Large Language Model (LLM) used for SaaS billing customer support.
The assistant is designed for a fictional company called PayFlow and must answer only using official billing policies.

The model is fine-tuned using LoRA (Low-Rank Adaptation) on top of a base instruction-tuned LLM.

---

2. Problem Statement

Large Language Models often produce confident but incorrect answers, known as hallucinations.
In billing and payments systems, hallucinations can cause:

• Financial loss
• Legal issues
• Loss of customer trust

This project addresses the question:

How can we constrain an LLM to answer strictly from approved billing policies and safely refuse when information is unavailable?

---

3. Why Hallucinations Are Dangerous in Billing Systems

Examples of real-world risks:

• Inventing refund policies
• Claiming unsupported payment methods
• Fabricating discounts or currencies

In enterprise systems, safe refusal is better than a wrong answer.

---

4. Project Goal

The goal of this project is to:

• Reduce hallucinations in billing-related questions
• Ensure strict policy adherence
• Teach the model when to refuse instead of guessing
• Quantify hallucination reduction using before/after evaluation

---

5. Model & Training Strategy

Base Model

• Instruction-tuned large language model (Mistral-style architecture)

Fine-Tuning Method

• LoRA (Low-Rank Adaptation)
• Only a small percentage of parameters are trained
• Base model weights remain unchanged

Why LoRA?

• Memory efficient
• Faster training
• Prevents catastrophic forgetting
• Industry-standard for alignment tasks

---

6. Dataset Design

The model is not trained directly on policy documents.

Instead:

• billing.md acts as the human source of truth
• Policies are manually converted into instruction–response pairs
• Only information present in the policy is allowed

Dataset Rules

• One concept → one behavior
• Explicit refusals for unknown information
• No assumptions or industry defaults

---

7. Policy-Driven Training Approach

The model is trained to follow this rule:

Answer only what is explicitly stated in PayFlow’s billing policy.
If information is missing, respond with a standard refusal.

Standard refusal phrase:

This information is not available in PayFlow’s billing policy.

This consistency is critical for hallucination control.

---

8. Training Process (Step-by-Step)

1. Environment setup in Google Colab
2. Load base model in 4-bit precision
3. Attach LoRA adapters
4. Load curated dataset (train.json)
5. Fine-tune LoRA adapters for 2–3 epochs
6. Save LoRA adapter weights only

No full model retraining is performed.

---

9. Evaluation Methodology

Evaluation is performed using:

• Known policy questions
• Edge cases
• Trap questions (questions not covered in policy)

Each response is classified as:

• ✅ Correct
• ⚠️ Safe but imperfect (over-refusal / verbosity)
• ❌ Hallucination (fabricated information)

---

10. Results Summary

• Hallucination rate before training: ~60–70%
• Hallucination rate after training: ~8–10%
• Approximate reduction: 75–90%

Detailed before/after comparisons are documented separately in RESULTS.md.

---

11. Challenges Faced

1. Partial Alignment Overconfidence

The model initially hallucinated more confidently after partial fine-tuning.

2. Over-Refusal

Excessive refusal occurred when refusal samples outweighed valid answers.

3. Dataset Contradictions

Similar questions with different expected behaviors caused instability.

Each issue was resolved through dataset normalization and retraining.

---

12. Key Learnings

• Hallucination reduction is a data problem, not a model problem
• Models hallucinate where policies are silent
• Over-refusal is safer than hallucination but must be balanced
• Alignment is iterative, not one-shot

---

13. Limitations & Future Improvements

• Some verbosity and response blending remains
• Further reduction (<5%) would require larger datasets
• Automated policy-to-dataset generation could improve scalability

---

14. Training Configuration Notes

LoRA adapters were successfully trained over 3 epochs, with training loss decreasing steadily, indicating effective policy alignment.

Training completed in 3 epochs with final training loss ≈ 1.14.

15. Conclusion

This project demonstrates a practical, enterprise-grade approach to hallucination control using LoRA.

Rather than chasing perfect answers, the model is trained to:

• Respect policy boundaries
• Avoid guessing
• Fail safely

This approach mirrors how real-world AI systems are deployed in billing and finance domains.

---