Qwen-1.7B-FNB-Service-Domain-Fine-Tuned

Qwen-1.7B-FNB-Service-Domain-Fine-Tuned

Model Overview

Qwen-1.7B-FNB-Service-Domain-Fine-Tuned is a domain-specialized large language model tailored for Food & Beverage (F&B) service and hospitality operations. It is designed to function as a practical service assistant and learning companion for F&B professionals, supervisors, and hospitality students.

The model follows a two-phase industry-aligned training strategy:

• Domain-Adaptive Pretraining (DAPT) on raw F&B service text
• Supervised Fine-Tuning (SFT) on curated instruction–response conversations

This approach strengthens domain reasoning, procedural accuracy, service-oriented communication, and hospitality-appropriate tone.

Base Model

• Base Model: Qwen-1.7B (Qwen3 family)
• Architecture: Decoder-only causal language model
• Parameters: ~1.7B
• Context Length: Up to 32k
• Serialization Format: safetensors
• Deployment Status: Fully merged, inference-ready

Training Methodology

Phase 1: Domain-Adaptive Pretraining (DAPT)

The base model was further pretrained on raw, unstructured F&B domain text from a public-domain F&B service textbook.

Objectives:

• Improve domain familiarity and vocabulary
• Reduce hallucinations in service scenarios
• Strengthen procedural grounding

Details:

• Raw text only (no instruction formatting)
• Continued causal language modeling
• QLoRA-based fine-tuning (4-bit) on NVIDIA T4 GPU

Phase 2: Supervised Fine-Tuning (SFT)

Instruction-tuning was performed using curated F&B service conversations, including:

• Guest complaint handling
• Allergy and food safety procedures
• Service recovery
• Upselling techniques
• Standard service sequences
• Operational opening/closing checklists

Objectives:

• Align responses with real-world service expectations
• Improve conversational and instructional usability
• Ensure hospitality-appropriate tone and clarity

Evaluation

Quantitative Evaluation

• Perplexity (held-out set): 7.02
  Indicates strong domain adaptation for structured operational content.

Qualitative Evaluation

Evaluated across seven F&B service scenarios:

• Cold food complaints
• Allergy disclosures
• Service recovery
• Upselling
• Operational checklists

Observed strengths:

• Clear step-by-step reasoning
• Professional hospitality tone
• Safety-aware recommendations
• Reduced generic/off-domain responses

Intended Use

• Food & Beverage service assistants
• Hospitality training platforms
• Hospitality education tools
• Operational guidance systems
• Conversational service agents

Limitations

• Domain-specialized; may underperform on unrelated technical/scientific tasks
• Not a replacement for certified food safety training
• Outputs should be reviewed by trained staff in real-world operations

Ethical Considerations

• Training data sourced from public-domain materials
• No personal or private data included
• Designed to support professionals, not replace human judgment

Citations

@misc{qwen3technicalreport, title={Qwen3 Technical Report}, author={Qwen Team}, year={2025}, eprint={2505.09388}, archivePrefix={arXiv}, primaryClass={cs.CL}, url={https://arxiv.org/abs/2505.09388} }

Fine Tuned model citation

@misc{til lu2026qwenfnb, title = {Qwen-1.7B-FNB-Service-Domain-Fine-Tuned}, author = {Tillu, Amey}, year = {2026}, howpublished = {\url{https://huggingface.co/Amey9766/Qwen-1.7B-FNB-Service-Domain-Fine-Tuned}}, note = {Domain-adapted and instruction-tuned Food & Beverage service model based on Qwen-1.7B} }

How to Use

Inference with Hugging Face Transformers

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model_id = "Amey9766/Qwen-1.7B-FNB-Service-Domain-Fine-Tuned"

tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    device_map="auto",
    torch_dtype=torch.float16,
    trust_remote_code=True
)

messages = [
    {"role": "system", "content": "You are an expert Food & Beverage service coach."},
    {"role": "user", "content": "A guest says their soup is cold. What should I do?"}
]

prompt = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True
)

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

output = model.generate(
    **inputs,
    max_new_tokens=200,
    temperature=0.7,
    top_p=0.9
)

print(tokenizer.decode(output[0], skip_special_tokens=True))