SkyAssist-Llama/README.md

---
license: mit
language:
- en

library_name: transformers

pipeline_tag: text-generation

base_model: meta-llama/Llama-3.2-3B-Instruct

tags:
- llama
- qlora
- peft
- airline
- customer-support
- conversational
- instruction-tuning
---

# ✈️ SkyAssist-Llama

SkyAssist-Llama is a domain-specific airline customer support large language model created by fine-tuning **Meta Llama 3.2 3B Instruct** using **QLoRA (Quantized Low-Rank Adaptation)**.

The model is designed to generate professional, empathetic, and context-aware responses for common airline customer support scenarios, including flight delays, cancellations, baggage inquiries, booking modifications, refunds, and check-in assistance.

---

# Model Details

| Property | Value |
|----------|-------|
| Base Model | Meta Llama 3.2 3B Instruct |
| Fine-tuning Method | QLoRA |
| Task | Airline Customer Support |
| Language | English |
| Architecture | Decoder-only Transformer |
| Training Objective | Supervised Fine-Tuning (SFT) |

---

# Training Data

The model was trained on **2,666** airline customer support conversations.

Rather than relying on an existing airline dataset, the training data was created through a multi-stage automated pipeline that transformed general customer support conversations into realistic airline-specific interactions.

The dataset generation process included:

- Data preprocessing
- LLM-based convertibility classification
- Airline domain transformation
- post-processing and cleaning
- JSON validation
- Sample-based manual review

The complete dataset is available on Hugging Face:

**Dataset:** https://huggingface.co/datasets/Lavanya177/airline-support-tone-clean

---

# Training Configuration

| Parameter | Value |
|-----------|-------|
| Base Model | Llama 3.2 3B Instruct |
| Fine-Tuning Method | QLoRA |
| LoRA Rank | 32 |
| LoRA Alpha | 64 |
| Learning Rate | 1e-3 |
| Epochs | 3 |
| Batch Size | 8 |
| Gradient Accumulation | 4 |
| Optimizer | AdamW |
| Scheduler | Cosine |
| Max Sequence Length | 2048 |
| Precision | BF16 |

---

# Evaluation

| Metric | Value |
|--------|------:|
| Training Loss | **0.4615** |
| Validation Loss | **0.5557** |
| Test Loss | **0.62** |

The model demonstrates good convergence while maintaining strong generalization on unseen airline customer support conversations.

---

# Intended Use

SkyAssist-Llama is intended for:

- Airline customer support assistants
- Conversational AI research
- Educational purposes
- Domain adaptation experiments
- LLM fine-tuning research

---

# Limitations

This model:

- does not access live airline booking systems
- cannot retrieve reservation details
- cannot process ticket changes or refunds
- may generate incorrect airline-specific policies
- should not replace official airline customer support

---

# Usage

## Load the model

```python
from transformers import AutoTokenizer, AutoModelForCausalLM

model_name = "Lavanya177/SkyAssist-Llama"

tokenizer = AutoTokenizer.from_pretrained(model_name)

model = AutoModelForCausalLM.from_pretrained(
    model_name,
    device_map="auto"
)
```
---

# Example

**System Prompt**

```
You are a professional airline customer support assistant...
```

**User**

```
My baggage has not arrived. What should I do?
```

**Assistant**

```
I'm sorry to hear that your baggage has not arrived. I understand how frustrating this situation can be.

Please provide your baggage claim number, flight number, travel date, and destination airport so I can assist you further. If you have already reported the issue at the airport, please also share your Property Irregularity Report (PIR) reference number.

Once these details are available, the airline can begin tracking your baggage and provide updates on its status.
```

---

# Citation

```bibtex
@misc{skyassistllama2026,
  title={SkyAssist-Llama: A Domain-Specific Airline Customer Support Large Language Model},
  author={Lavanya Singh},
  year={2026}
}
```

# License

This project is released under the MIT License.