DroidSenior-7B

DroidSenior-7B is a 7B coding assistant specialized in Android development, fine-tuned from Qwen2.5-Coder-7B-Instruct. It aims to answer like a senior/staff Android engineer: modern idiomatic Kotlin, Jetpack Compose, Coroutines/Flow, Hilt and Room, the official Android architecture guidance, explicit trade-offs, no deprecated APIs and tests where they add value. It answers in English and Turkish.

• Weights for transformers / vLLM: this repo
• GGUF for Ollama / LM Studio / llama.cpp: celebibasar/DroidSenior-7B-GGUF

Recommended system prompt

The model was trained with this system prompt on its highest-quality data. Use it for best results:

You are DroidSenior, a staff-level Android engineer. Write idiomatic, modern Kotlin (Jetpack Compose, Coroutines/Flow, Hilt, Room) following the official Android architecture guidance. Explain key design decisions and trade-offs briefly, never use deprecated APIs, avoid over-engineering, and include tests when they add value. Answer in the user's language.

Quick start

transformers

from transformers import AutoModelForCausalLM, AutoTokenizer

repo = "celebibasar/DroidSenior-7B"
SYSTEM_PROMPT = """You are DroidSenior, a staff-level Android engineer. Write idiomatic, modern Kotlin (Jetpack Compose, Coroutines/Flow, Hilt, Room) following the official Android architecture guidance. Explain key design decisions and trade-offs briefly, never use deprecated APIs, avoid over-engineering, and include tests when they add value. Answer in the user's language."""

tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(repo, torch_dtype="auto", device_map="auto")

messages = [
    {"role": "system", "content": SYSTEM_PROMPT},
    {"role": "user", "content": "Implement a paginated list screen with Compose, Room and Paging 3."},
]
text = tok.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tok(text, return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=1500, do_sample=True, temperature=0.3, top_p=0.9)
print(tok.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

vLLM (OpenAI-compatible server)

vllm serve celebibasar/DroidSenior-7B --max-model-len 16384

Ollama

ollama run hf.co/celebibasar/DroidSenior-7B-GGUF:Q4_K_M

For the recommended system prompt, download Modelfile and the GGUF file from the GGUF repo and run ollama create droid-senior -f Modelfile.

Training

1. Open data: Android/Kotlin examples from public Hugging Face datasets, filtered for Android relevance, Java-only answers, deprecated APIs (AsyncTask, GlobalScope, kapt, synthetics, ...) and near-duplicates.
2. Senior distillation: a taxonomy of 12 areas x 9 task types (feature implementation, code review, legacy refactoring, architecture decisions, debugging, testing, performance, deep dives, system design). A teacher model (Qwen/Qwen3-Coder-30B-A3B-Instruct-FP8) generated realistic questions and answered them with a detailed staff-engineer prompt.
3. Judge filtering: every answer scored on correctness, modernity, architecture, code quality, completeness and clarity; only high-scoring answers were kept.
4. SFT: LoRA (bf16) on all attention and MLP projections, loss on assistant tokens only. Senior data used the system prompt above; open data used a generic prompt (quality conditioning).
5. DPO: on-policy preference tuning — the SFT model sampled 4 answers per prompt, the judge picked the best/worst (654 pairs, beta=0.1, lr=5e-06, LoRA r=32).

Data

source	examples	notes
Synthetic senior Q&A distilled from Qwen/Qwen3-Coder-30B-A3B-Instruct-FP8	4143	judge-filtered (kept 4143 of 8788)
JetBrains/KExercises	4000	filtered: Android/Kotlin, legacy APIs, dedup; license: CC-BY-NC-SA-4.0 (non-commercial)
giggiovpg/android-kotlin-compose-compiler-verified	5767	filtered: Android/Kotlin, legacy APIs, dedup
giggiovpg/ornith-android-instruct	5176	filtered: Android/Kotlin, legacy APIs, dedup
glaiveai/glaive-code-assistant-v3	586	filtered: Android/Kotlin, legacy APIs, dedup
ise-uiuc/Magicoder-OSS-Instruct-75K	33	filtered: Android/Kotlin, legacy APIs, dedup
m-a-p/CodeFeedback-Filtered-Instruction	49	filtered: Android/Kotlin, legacy APIs, dedup
microsoft/NextCoderDataset	5000	filtered: Android/Kotlin, legacy APIs, dedup
theblackcat102/evol-codealpaca-v1	1566	filtered: Android/Kotlin, legacy APIs, dedup

Hyperparameters

stage	method	lora_r	lora_alpha	epochs	learning_rate	effective_batch	max_seq_len	train_examples
SFT	LoRA (bf16)	64	64	2	0.0001	32	8192	7793
DPO	LoRA	32	32	1	5e-06	16	8192	654

Evaluation

Evaluated on 150 held-out questions (never used in training), scored 1-10 by Qwen/Qwen3-Coder-30B-A3B-Instruct-FP8 with the same rubric for every model.

model	correctness	modernity	architecture	code_quality	completeness	clarity	overall	legacy_api_rate
base	6.91	6.13	5.92	6.09	5.56	6.6	5.84	0.09
sft	8.08	7.47	7.66	7.46	7.05	7.87	7.23	0.0
dpo	8.13	7.54	7.73	7.53	7.08	8.02	7.29	0.01

Head-to-head vs. the base model (overall score per question):

model	win	tie	loss
sft	76%	17%	7%
dpo	73%	23%	5%

Caveats: LLM-as-judge scores, judged by the same model family that generated the training data, on questions drawn from the same generator distribution. Treat them as a relative signal, not an absolute benchmark.

Limitations

• 7B model: it can still produce code that does not compile or uses APIs incorrectly. Review generated code.
• Knowledge of libraries is limited to what the base model and training data contain; very recent APIs may be missing.
• Training data is largely synthetic and judged by an LLM; it inherits the teacher's biases and mistakes.

License

Released under apache-2.0. Base model: Qwen2.5-Coder-7B-Instruct (Apache-2.0). Training data includes non-commercially licensed sets (JetBrains/KExercises); check their terms before commercial use.