This model is a fine-tuned and fully merged variant of Meta LLaMA-3.2-1B-Instruct, adapted for hospitality revenue management and decision-support tasks. Fine-tuning was performed using supervised fine-tuning (SFT) with LoRA adapters to efficiently specialize the base model while keeping training costs manageable. The training data consisted of domain-specific instruction–response pairs covering pricing strategy, demand forecasting reasoning, inventory and capacity trade-offs, seasonality impacts, and revenue optimization scenarios commonly encountered in hospitality operations. Training was conducted using the Hugging Face Trainer framework with mixed-precision (FP16), gradient accumulation, and an 8-bit AdamW optimizer to support memory-efficient training. After convergence, the LoRA adapters were merged into the base model weights, producing a standalone model that no longer requires external adapters and can be deployed directly for inference in production or Hugging Face Spaces.
license: llama3
base_model: meta-llama/Llama-3.2-1B-Instruct
library_name: transformers
pipeline_tag: text-generation
tags:
- llama
- llama-3
- instruction-tuned
- hospitality
- revenue-management
- decision-support
- merged
- safetensors
model_type: causal-lm
language:
- en
LLaMA-3.2-1B-Instruct-Revenue-Decision (Merged)
Model Summary
LLaMA-3.2-3B-Instruct-Revenue-Decision-Merged is a domain-adapted large language model designed to assist with hotel revenue management and decision-making tasks.
The model is based on Meta LLaMA-3.2-1B-Instruct and was fine-tuned using supervised fine-tuning (SFT) on hospitality-specific data, followed by LoRA adaptation and a full weight merge into the base model.
This repository contains the fully merged model weights, meaning no LoRA adapters are required at inference time.
Intended Use
This model is intended for decision support and analytical reasoning in hospitality and travel contexts, including:
- Hotel pricing and rate strategy recommendations
- Occupancy and demand-based decision reasoning
- Revenue optimization scenarios
- Business and operational trade-off analysis
- Scenario-based “what should I do?” reasoning
It is suitable for:
- Research
- Prototyping
- AI agents
- Internal decision-support tools
Not Intended For
- Medical, legal, or financial advice
- Fully autonomous decision-making without human oversight
- High-risk or safety-critical systems
Training Details
- Base Model:
meta-llama/Llama-3.2-1B-Instruct
- Fine-Tuning Method: Supervised Fine-Tuning (SFT)
- Adaptation Technique: LoRA
- LoRA Rank (r): 16
- LoRA Alpha: 32
- Final Model: Fully merged LoRA + base weights
- Framework: Hugging Face Transformers + PEFT
Model Architecture
- Decoder-only causal language model
- Transformer architecture
- Chat-optimized instruction format
- Supports chat templates via
chat_template.jinja
Files Included
This repository includes all files required for inference and deployment:
model-00001-of-00002.safetensorsmodel-00002-of-00002.safetensorsmodel.safetensors.index.jsonconfig.jsongeneration_config.jsontokenizer.jsontokenizer_config.jsonspecial_tokens_map.jsonchat_template.jinja
Base model citation-
@misc{llama3,
title = {The LLaMA 3 Model Family},
author = {Meta AI},
year = {2024},
howpublished = {\url{
https://ai.meta.com/llama/}}
}
Tillu, A. (2026). LLaMA-3.2-3B-Instruct Revenue Decision Model (Merged). Hugging Face.
https://huggingface.co/Amey9766/llama-3.2-3b-instruct-revenue-decision-mergedAuthor-
Amey Tillu
Hospitality Data Analyst AI/ML Hobbyist
Usage Example
1from transformers import AutoTokenizer, AutoModelForCausalLM
2import torch
3
4model_id = "Amey9766/llama-3.2-3b-instruct-revenue-decision-merged"
5
6tokenizer = AutoTokenizer.from_pretrained(model_id)
7model = AutoModelForCausalLM.from_pretrained(
8 model_id,
9 torch_dtype=torch.float16,
10 device_map="auto"
11)
12
13prompt = "Demand is increasing but weekday occupancy is low. What pricing strategy should a hotel consider?"
14
15inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
16
17with torch.no_grad():
18 outputs = model.generate(
19 **inputs,
20 max_new_tokens=150,
21 temperature=0.7
22 )
23
24print(tokenizer.decode(outputs[0], skip_special_tokens=True))
25
26
