| --- |
| base_model: unsloth/gemma-4-E4B-it-unsloth-bnb-4bit |
| library_name: peft |
| pipeline_tag: text-generation |
| language: |
| - en |
| tags: |
| - base_model:adapter:unsloth/gemma-4-E4B-it-unsloth-bnb-4bit |
| - lora |
| - sft |
| - transformers |
| - trl |
| - unsloth |
| - gemma |
| - android |
| - adb |
| - function-calling |
| - tool-use |
| --- |
| |
| # DroidADB: Gemma 4 E4B LoRA for Android/ADB Tool Calling |
|
|
| DroidADB is a LoRA adapter for Gemma 4 E4B Instruct. It chats normally, and when a request involves Android or ADB it calls an `adb` function tool instead of just printing a command. This is a personal project, trained on a personal Google Colab T4 GPU. It isn't affiliated with or funded by any organization. |
|
|
| ## Model Details |
|
|
| ### Model Description |
|
|
| This adapter was trained with supervised fine-tuning (SFT) using Unsloth and TRL. It uses Gemma 4's native function-calling format. Each training conversation has this shape: |
|
|
| user request → `adb` tool call → tool result → final assistant reply |
|
|
| A small set of general chat examples was mixed in so the model can still hold a normal conversation. |
|
|
| - **Developed by:** Muneesh Kanimbakam |
| - **Funded by:** Self-funded (personal project) |
| - **Shared by:** Muneesh Kanimbakam |
| - **Model type:** Causal LM, LoRA adapter (PEFT), function-calling / tool-use |
| - **Language(s):** English |
| - **License:** <TODO: must match the base Gemma 4 license> |
| - **Finetuned from model:** `unsloth/gemma-4-E4B-it`, loaded in 4-bit (resolved to `unsloth/gemma-4-E4B-it-unsloth-bnb-4bit`) |
|
|
| ### Model Sources |
|
|
| - **Repository:** https://huggingface.co/<HF_USERNAME>/<MODEL_NAME> |
|
|
| ## Uses |
|
|
| ### Direct Use |
|
|
| - Turning natural-language Android requests into `adb` tool calls, e.g. "list installed packages", "show device logs", "get the battery status" |
| - General conversation, such as simple explanations, planning and summaries |
| - Serving as the model in a host app's tool loop: the host runs the tool and the model answers using the tool result |
|
|
| ### Downstream Use |
|
|
| - Plugging into any agent framework that supports OpenAI/Gemma-style function tools |
| - Adding more tools through a tool registry (`register_tool`) without retraining |
| - Exporting to GGUF (`q4_k_m`) for llama.cpp or Ollama |
|
|
| ### Out-of-Scope Use |
|
|
| - Running ADB commands automatically without a human approving them |
| - Rooting a device, modifying system partitions, or bypassing device security |
| - Production device management, or any high-stakes or safety-critical use |
| - Non-Android technical tasks that need expert accuracy |
|
|
| ## Bias, Risks, and Limitations |
|
|
| - **Risk of wrong commands:** the model may produce ADB commands that are incorrect, outdated, or harmful. Always review a command before running it. |
| - **The adapter itself doesn't execute anything.** Safety comes from the host runtime. The reference runtime in the notebook: |
| - dry-runs by default (`ALLOW_COMMAND_EXECUTION = False`) |
| - accepts only commands that start with `adb` |
| - blocks shell operators (`|`, `;`, `&&`, `||`, `>`, `>>`) and runs without `shell=True` |
| - asks for confirmation before risky commands (`install`, `uninstall`, `rm`, `mv`, `reboot`, `settings put`, `chmod`, `chown`, `su`, `wipe`, `format`, `pm clear`) |
| - **Synthetic tool results:** in training, every tool result was a placeholder that said the command was not executed. The model never saw real device output, so it may handle real outputs poorly. |
| - **Limited general chat:** only 4 general-chat examples were included, so conversational quality relies mostly on the base model. |
| - Inherits the biases and limitations of the base Gemma model. 4-bit quantization may slightly reduce quality. |
|
|
| ### Recommendations |
|
|
| Keep execution turned off by default. Require a human to approve each command, and pin the target device (`ADB_SERIAL`). Prefer read-only diagnostic commands. |
|
|
| ## How to Get Started with the Model |
|
|
| ```python |
| from unsloth import FastModel |
| from unsloth.chat_templates import get_chat_template |
| |
| model, tokenizer = FastModel.from_pretrained( |
| model_name=""Muneeshk/gemma-4-adb-agent"", |
| max_seq_length=4096, |
| load_in_4bit=True, |
| ) |
| tokenizer = get_chat_template(tokenizer, chat_template="gemma-4") |
| FastModel.for_inference(model) |
| |
| ADB_TOOL = { |
| "type": "function", |
| "function": { |
| "name": "adb", |
| "description": "Run one Android Debug Bridge command on an explicitly selected device.", |
| "parameters": { |
| "type": "object", |
| "properties": { |
| "command": {"type": "string", "description": "Full command beginning with adb. Never include a shell pipeline."}, |
| "confirm": {"type": "boolean", "description": "Set true only after user explicitly confirms a risky command."}, |
| }, |
| "required": ["command"], |
| }, |
| }, |
| } |
| |
| messages = [ |
| {"role": "system", "content": "You are DroidADB, a friendly general-purpose assistant with strong Android and ADB expertise."}, |
| {"role": "user", "content": "How do I list all installed Android packages?"}, |
| ] |
| inputs = tokenizer.apply_chat_template( |
| messages, tools=[ADB_TOOL], tokenize=True, add_generation_prompt=True, |
| return_dict=True, return_tensors="pt", |
| ).to(model.device) |
| |
| out = model.generate(**inputs, max_new_tokens=256, do_sa**le=True, |
| tem**rature=1.0, top_p=0.95, top_k=64,**se_cache=True) |
| print(tokenizer.de**de(out[0][inputs["input_ids"].shape[-1\]:], skip_special_tokens=Fals**) |
