---
language:
- en
- es
- code
tags:
- code-generation
- android
- kotlin
- java
- jetpack-compose
- qwen2.5
- unsloth
- gguf
- ollama
- 3b
license: apache-2.0
base_model: Qwen/Qwen2.5-3B-Instruct
datasets:
- giggiovpg/ornith-android-instruct
- giggiovpg/android-kotlin-compose-compiler-verified
- microsoft/NextCoderDataset
- glaiveai/glaive-code-assistant-v3
---

# 🤖 LaboAI-0.3.3-3B

This is a lightweight yet capable language model (3B parameters) fine-tuned specifically for generating, understanding, and debugging **Kotlin** code and **Android** development (with a strong emphasis on Jetpack Compose, Coroutines, and modern architectures). 

This version (0.3.3) applies the same proven, multi-dataset training recipe as the 1.5B version, but scaled up to the 3B architecture for improved reasoning, better context understanding, and more reliable code generation. It is optimized using **QLoRA (4-bit)** to run efficiently on consumer hardware (e.g., NVIDIA RTX 3060 12GB, or RTX 4060).

## 📋 Model Details

- **Developed by:** Mmxa
- **Organization:** LaboAI
- **Model type:** Causal Language Model (Code Generation)
- **Languages:** Kotlin, Java, English, Spanish (instructions)
- **License:** Apache 2.0 (inherited from Qwen2.5)
- **Base model:** [Qwen/Qwen2.5-3B-Instruct](https://huggingface.co/Qwen/Qwen2.5-3B-Instruct)

## 🚀 Uses

### Direct Use
- Generating robust boilerplate for Activities, Fragments, ViewModels, and Repositories in Kotlin.
- Creating complex modern UI components with **Jetpack Compose**.
- Debugging compilation errors or logic flaws in Android code snippets.
- Translating legacy Java logic into modern, idiomatic Kotlin.

### Ecosystem Use (Recommended)
This model shines when used as a local coding assistant via **Ollama** and the **Continue** extension in VS Code. This guarantees complete privacy (your code never leaves your machine) and low latency.

### Out-of-Scope Uses
- It is not optimized for general chat, creative writing, or complex mathematical reasoning.
- It should not be used to generate malicious code or exploits.
- All generated code must be reviewed by a human developer before being merged into a main branch.

## ⚠️ Limitations and Risks
- **API Hallucinations:** In rare cases, it might suggest deprecated Android APIs instead of modern alternatives.
- **Context Window:** Optimized for 1024 tokens during training. It is not suitable for analyzing massive, multi-thousand-line codebase files all at once.
- **Dependencies:** It does not have real-time knowledge of the latest Android library updates.

## 💻 How to Get Started (Local Setup)

This repository includes both the original format (`safetensors`) and the quantized format (`GGUF` Q4_K_M). 

1. Install [Ollama](https://ollama.com/).
2. Run the model directly from Hugging Face:
   ```bash
   ollama run hf.co/LaboAI/LaboAI-0.3.3-3B:Q4_K_M