imend35/Begonia-AI-GGUF

Begonia AI — GGUF Models

This repository contains GGUF exports of the fine-tuned Begonia AI merged model.

Available Files

File	Approximate Size
Begonia-AI-1.5B-F16.gguf	2.88 GiB
Begonia-AI-1.5B-Q4_K_M.gguf	0.92 GiB

The Q4_K_M file is approximately 68.1% smaller than F16.

The Q4_K_M version passed a basic text-generation smoke test using llama.cpp. This test confirms execution, not comprehensive answer quality.

Project Overview

Begonia AI is a Turkish-language hotel reception assistant developed for the fictional Begonvil Hotel as an educational capstone project.

The project uses Qwen2.5-1.5B-Instruct as its base model and applies QLoRA fine-tuning using synthetic hotel reception conversations.

The assistant is designed to answer hotel-related questions, retrieve information from an official hotel guide using Chroma-based RAG, and execute structured tools for room information and accommodation costs.

Training

• Base model: Qwen/Qwen2.5-1.5B-Instruct
• Fine-tuning: QLoRA, 4-bit NF4
• LoRA rank: 16
• LoRA alpha: 32
• LoRA dropout: 0.05
• Epochs: 3
• Training examples: 450
• Validation examples: 60
• Held-out evaluation examples: 60

Automatic Evaluation

Metric	Before	After
Identity	0%	100%
Guide-related response criterion	10%	65%
Agent JSON format	90%	100%
Agent tool selection	10%	95%

These are project-specific automated checks, not general-purpose model accuracy scores or independent benchmarks.

Additional Observations

Two tested agent requests selected the expected tools and returned the verified tool outputs. Both received a score of 1.0 from the Gemini-based evaluation used in this project.

The initial RAG evaluation retrieved relevant source passages but the generated answers were not consistently correct. A later Gradio interface test correctly reported check-in and check-out times, although an additional unverified detail appeared.

The GGUF Q4_K_M version successfully generated Turkish text using llama.cpp, but some grammatical issues and repetition of system instructions were observed.

Limitations

• This is an educational prototype, not a production booking system.
• The hotel and its training records are fictional.
• The model can hallucinate prices, policies, and other details.
• Retrieval success does not guarantee answer correctness.
• Tool use and numerical outputs should be independently verified.
• No real reservation is created by this assistant.
• Human review is required before operational deployment.

Related Repositories

• Merged model
• LoRA adapter
• Dataset
• GGUF