phi3-hospitality-support-bot

🏨 Phi-3 Hospitality Support Bot Fine-Tuned Microsoft Phi-3 Mini 4K (Merged, Full Model)

Author: Amey Tillu Model Size: 3B parameters (merged LoRA) Use Case: Hotel Concierge AI, Guest Support, Hospitality Automation

📌 Overview

This model is a domain-specialized version of Microsoft's Phi-3 Mini 4K Instruct, fine-tuned extensively on custom hospitality & hotel management datasets created to power an AI Concierge / Support Bot.

It has been trained to understand and respond to:

Guest questions

Reservation-related queries

Amenities & services

Hotel policies

Parking, dining, spa, pool, transportation

Housekeeping, rooms, check-in, check-out

Local attractions & recommendations

The goal is to deploy this model inside a Hotel AI Agent capable of handling real-time guest conversations across web, mobile, and phone.

🧠 Base Model

Name: microsoft/Phi-3-mini-4k-instruct

Type: Small Language Model (SLM)

Context Length: 4,000 tokens

License: MIT License

🔧 Fine-Tuning Details

Method: QLoRA (4-bit NF4 quantization)

Training Strategy: Supervised Fine-Tuning (SFT)

Epochs: 3

Batch Size: 2

Gradient Accumulation: 4

Learning Rate: 2e-4

Sequence Length: 2048 tokens

GPU Used: NVIDIA A100 (High-RAM)

Frameworks:

transformers

trl

peft

bitsandbytes

The LoRA adapters were merged into the base model using merge_and_unload(), resulting in a fully standalone fine-tuned model.

📚 Dataset

The model was fine-tuned using a large, custom JSONL dataset containing over 50,000 high-quality hospitality-domain instruction-response pairs, covering:

⭐ Guest FAQs

(check-in times, breakfast, parking, WiFi, pool, spa, etc.)

⭐ Rooms & Services

(room types, housekeeping, amenities, upgrades)

⭐ Policies

(refund, pets, ID requirements, smoking, payment methods)

⭐ F&B

(restaurant menu, timing, bar, breakfast options)

⭐ Spa, Gym, Pool

(timings, charges, facilities)

⭐ Transportation

(airport pickup, shuttle, taxi, metro, parking)

⭐ Local Recommendations

(restaurants, tourist spots, things to do)

The dataset was crafted manually + semi-synthetically to represent real hotel guest conversations.

Evaluation

After fine-tuning, the model was evaluated on:

✔ Perplexity

Perplexity: 1.28