LaboAI-0.4.0-3B is a 3B-parameter code model fine-tuned for Android / Kotlin / Jetpack Compose development. It is built on top of Qwen2.5-Coder-3B-Instruct using QLoRA with Unsloth, and the LoRA weights are merged into the released model.
Model details
Developer
LaboAI
Base model
Qwen/Qwen2.5-Coder-3B-Instruct
Parameters
~3B
Fine-tuning method
QLoRA (4-bit base, LoRA r=32, alpha=64)
Context used in training
1024 tokens
Languages
Kotlin, Java, general code; English and Spanish prompts
Available formats
Merged safetensors, GGUF q4_k_m
Intended use
Generating and explaining Kotlin code for Android apps
Jetpack Compose UI snippets, ViewModels, state handling, Room, Retrofit, coroutines
Fixing and refactoring small to medium Kotlin snippets
Local inference on modest hardware via the GGUF build (llama.cpp, Ollama, LM Studio)
Prompt format
The model was trained with a plain instruction/response template (not the Qwen chat template), so for best results use:
### Instruction:
{your request}
### Response:
How to use
Transformers
fromtransformersimportAutoTokenizer,AutoModelForCausalLMimporttorchrepo="LaboAI/LaboAI-0.4.0-3B"tokenizer=AutoTokenizer.from_pretrained(repo)model=AutoModelForCausalLM.from_pretrained(repo,torch_dtype=torch.float16,device_map="auto")prompt=("### Instruction:\n""Write a Jetpack Compose screen with a counter and a button to increment it.\n\n""### Response:\n")inputs=tokenizer(prompt,return_tensors="pt").to(model.device)out=model.generate(**inputs,max_new_tokens=512,temperature=0.2,do_sample=True)print(tokenizer.decode(out[0],skip_special_tokens=True))
llama.cpp / GGUF
llama-cli -m LaboAI-0.4.0-3B-Q4_K_M.gguf -c 2048 --temp 0.2 \
-p "### Instruction:\nCreate a Room DAO for a Note entity.\n\n### Response:\n"
Training data
The training mix was filtered for Android/Kotlin/Jetpack content (keyword filter on the text fields) and formatted into the instruction/response template above:
Examples with empty or near-empty text were removed, and the final dataset was shuffled (seed 42).
Training procedure
Hyperparameter
Value
LoRA rank / alpha / dropout
32 / 64 / 0
Target modules
q, k, v, o, gate, up, down projections
Max sequence length
1024
Per-device batch size
2
Gradient accumulation
4 (effective batch 8)
Steps
4,500
Warmup steps
225
Learning rate
1e-4 (cosine)
Weight decay
0.01
Optimizer
AdamW 8-bit
Precision
fp16 (T4)
Hardware
1x NVIDIA T4 (Kaggle)
Framework
Unsloth + TRL SFTTrainer
Limitations and risks
It is a small 3B model: it can produce code that does not compile, uses deprecated or non-existent APIs, or contains subtle bugs. Always review and test the output.
Trained with a 1024-token context, so very long files or multi-file projects may degrade quality.
The keyword-based filtering may have let in some off-topic examples.
Android and Compose APIs change quickly; the model may not know the latest library versions.
Not evaluated on formal benchmarks yet; no quantitative results are claimed.
License
This model is derived from Qwen2.5-Coder-3B-Instruct and inherits its license (Qwen Research License). Please check the base model's license for terms of use, especially for commercial use. The training datasets have their own licenses, which you should also review.
