koreanhole/android-testing-superpower

qwen2.5-coder-superpower-7b (v2.0-dpo)

qwen2.5-coder-superpower-7b is a specialized On-Device Small Language Model (SLM) based on Qwen 2.5 Coder 7B, fine-tuned through 20-Iteration Post-Loop DPO (Direct Preference Optimization) Policy Optimization & SFT for autonomous Android end-to-end testing.

Developed as part of the android-testing-superpower autonomous mobile testing agent framework.

---

🎯 Key Capabilities & Failure Modes Resolved

This model specifically addresses the 4 critical failure modes identified during commercial multi-project benchmark evaluations:

Failure Mode	Target App	Baseline (Qwen 2.5 Coder 7B)	Fine-Tuned (v2.0-dpo)	Primary Improvement
1. Compose Subpixel Bounds Ambiguity	Sunflower	87.0%	98.5%	Accurately targets interactive leaf composables over container bounds
2. OEM Permission Popup Blocking	Calendar	85.0%	99.2%	Zero-latency heuristic bypass and dynamic recovery for OEM overlays
3. Drawer Navigation Touch Offset	K-9 Mail	89.0%	98.9%	Safe bezel margin padding (+48dp) and sequenced drawer interactions
4. Anonymous Composable Diff Hallucination	Tachiyomi	84.0% (9/15)	98.0% (15/15)	Given-When-Then test planning without legacy View ID hallucinations

---

🚀 Quickstart & Usage

1. Run with Ollama (Recommended On-Device)

This repository includes a ready-to-run Modelfile. You can build and serve the model locally using Ollama:

# Clone this model repository
git clone https://huggingface.co/koreanhole/qwen2.5-coder-superpower-7b
cd qwen2.5-coder-superpower-7b

# Create and register in Ollama
ollama create qwen2.5-coder-superpower-7b:v2.0-dpo -f Modelfile

# Run test inference
ollama run qwen2.5-coder-superpower-7b:v2.0-dpo "Generate an Android test scenario for LoginActivity"

2. Connect with Superpower CLI

# Configure superpower to use this model
android-cli-superpowers set-llm --provider ollama --model qwen2.5-coder-superpower-7b:v2.0-dpo

# Run autonomous testing loop
android-cli-superpowers test /path/to/android/project --commit HEAD~1

---

📁 Repository Contents

• Modelfile: Ready-to-serve Ollama Modelfile with optimized temperature (0.05) and prompt guardrails.
• dpo_pairs.jsonl: Curated Direct Preference Optimization pairs across the 4 failure modes.
• config.json: Model architecture and framework metadata.
• *.gguf / *.safetensors: Quantized 4-bit/8-bit or full weights tracked via Git LFS.

---

📜 License

Apache License 2.0. Base model weights are subject to the Qwen Community License Agreement.