antiableofnormies/qwen3.5-9b-lora-android-dev

Code Style LoRA for Qwen3.5-9B

A QLoRA adapter fine-tuned on 201 Android project source files to teach Qwen3.5-9B a specific coding style. Designed for llama.cpp (via GGUF conversion) or HuggingFace Transformers + PEFT.

What this adapter does

• Learns Kotlin coding patterns: naming conventions, import ordering, brace style, comment density
• Captures Android architecture patterns: ViewModel, Repository, Room DB, Retrofit networking
• Embeds dependency injection style (Hilt)
• No unwanted boilerplate — trained to generate exactly the level of verbosity you use

Training details

Parameter	Value
Base model	Qwen/Qwen3.5-9B
Method	QLoRA (4-bit NF4 via Unsloth)
LoRA rank	r=16, alpha=16
Target modules	q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
Training data	201 source files from 2 Android projects
File types	Kotlin (169), Gradle KTS (18), XML resources (14)
Epochs	3
Sequence length	2048
Learning rate	2e-4 (cosine scheduler)
Hardware	T4 GPU (16GB VRAM) via Google Colab
Framework	Unsloth + TRL SFTTrainer

Usage with Transformers + PEFT

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

model = PeftModel.from_pretrained(
    AutoModelForCausalLM.from_pretrained("Qwen/Qwen3.5-9B", device_map="auto"),
    "antiableofnormies/qwen3.5-9b-lora-android-dev",
)
tokenizer = AutoTokenizer.from_pretrained("antiableofnormies/qwen3.5-9b-lora-android-dev")

Usage with llama.cpp

python convert_lora_to_gguf.py     --base Qwen3.5-9B-Q4_K_M.gguf     --lora ./qwen3.5-style-lora/     --output style-adapter.gguf

llama-server -m Qwen3.5-9B-Q4_K_M.gguf     --lora style-adapter.gguf     --host 0.0.0.0 -ngl 99 --ctx-size 32768 --port 8080 --mlock

# Optional: merge into a standalone GGUF
llama-export-lora -m Qwen3.5-9B-Q4_K_M.gguf     --lora style-adapter.gguf     -o qwen3.5-code-style-merged.gguf

Intended use

This adapter is designed for a local coding assistant (opencode + llama.cpp) that:

• Runs entirely on your machine (no API calls)
• Respects your existing code conventions
• Avoids unwanted boilerplate that cloud APIs tend to add
• Matches the patterns found in the training projects

Limitations

• Trained on a personal codebase — style may not generalize to unrelated projects
• LoRA rank 16 captures high-level style (naming, structure) but not deep domain knowledge
• Jetpack Compose UI code was explicitly excluded from training
• English-only code and comments