This repository contains a domain-specific fine-tuned LLaMA-3 8B model for generating TRObject code from natural language prompts.
TRObject is used in the Clomosy mobile application development platform. The goal of this model is to help users generate copy-ready TRObject code snippets for common mobile UI and application-development tasks.
Project Motivation
General-purpose LLMs do not reliably understand TRObject syntax, Clomosy-specific UI patterns, or platform-specific development workflows.
To address this gap, I created a custom instruction-style dataset and fine-tuned a LLaMA-3 8B model for TRObject code generation.
The project focuses on:
Natural language to TRObject code generation
Mobile UI component generation
Form, button, input, event, and database-related code snippets
Domain-specific adaptation of a general-purpose LLM
Base Model and Fine-Tuning
Base model:unsloth/llama-3-8b-bnb-4bit
Architecture: LLaMA-3 8B
Fine-tuning method: Supervised fine-tuning with LoRA adapters
Optimization: 4-bit quantization with Unsloth
Libraries: Hugging Face Transformers, TRL, Unsloth, PEFT/LoRA
Dataset
The model was fine-tuned on a custom instruction-style dataset containing natural language prompts and corresponding TRObject code outputs.
Educational use in custom code-generation workflows
Assisting developers with repetitive Clomosy/TRObject UI patterns
Limitations
This model is experimental and should not be used as a fully autonomous code generator without human review.
Known limitations:
May generate syntactically incorrect TRObject code for complex requests
May hallucinate unsupported components or APIs
Works better on patterns similar to the training data
Requires developer validation before use in production
Dataset size is limited, so generalization may be weak outside known TRObject patterns
Recommended Usage
Use the model as an assistant for generating draft TRObject snippets. All generated code should be manually reviewed and tested inside the Clomosy/TRObject development environment.
Data Safety Note
The dataset and examples should only include sanitized, non-sensitive, and non-confidential code samples.
Do not use private credentials, production data, customer data, or proprietary business logic in public examples.
Developer
Developed by Mehmet Öden as part of an AI/LLM internship project focused on domain-specific code generation for TRObject and Clomosy-style mobile application development.
