---
base_model: google/gemma-4-E2B-it
library_name: transformers
license: apache-2.0
pipeline_tag: text-generation
tags:
- gemma4
- gguf
- tool-use
- privacy
- on-device-ai
- lora
- unsloth
- quantized
language:
- en
datasets:
- custom
model-index:
- name: snowclaw-gemma4-e2b-ft
  results:
  - task:
      type: text-generation
      name: Tool Use Accuracy
    metrics:
    - type: accuracy
      value: 100
      name: Tool Use (7/7)
---

# SnowClaw — Fine-tuned Gemma 4 E2B for Privacy-First Tool Use

A fine-tuned [Gemma 4 E2B](https://ai.google.dev/gemma) model optimized for **on-device AI tool use** in the SnowClaw desktop agent. Achieves **100% tool use accuracy (7/7)** on our evaluation set.

## Model Details

| Property | Value |
|----------|-------|
| **Base Model** | `google/gemma-4-E2B-it` (Gemma 4 E2B Instruct) |
| **Method** | LoRA (rank=64, alpha=64) via [Unsloth](https://github.com/unslothai/unsloth) |
| **Training Data** | 2,000 synthetic tool-use examples |
| **Epochs** | 15 |
| **Final Loss** | 0.040 |
| **Hardware** | NVIDIA RTX 3090 (24GB VRAM) |
| **Quantization** | Q4_K_M (GGUF) |
| **Vision** | Multimodal — includes mmproj for image understanding |

## Files

| File | Size | Description |
|------|------|-------------|
| `gemma-4-e2b-it.Q4_K_M.gguf` | 3.2 GB | Main model (Q4_K_M quantized) |
| `gemma-4-e2b-it.BF16-mmproj.gguf` | 942 MB | Vision projector (multimodal image encoder) |
| `Modelfile` | 205 B | Ollama registration file |

## Intended Use

SnowClaw is a **privacy-first desktop AI agent** that runs entirely on-device. This model is fine-tuned for:

- **Tool Use**: Executing system commands, browsing files, managing contacts/calendar
- **Code Generation**: Writing and executing Python/AppleScript in a sandboxed environment
- **Screenshot Analysis**: Understanding screen content via vision capabilities
- **Privacy**: All processing stays local — zero data leaves the device

## How to Use

### With Ollama

```bash
# Download the GGUF files, then register with Ollama:
ollama create snowclaw -f Modelfile

# Run
ollama run snowclaw
```

### With llama.cpp

```bash
# Text only
./llama-cli -m gemma-4-e2b-it.Q4_K_M.gguf -p "List files in my Downloads folder"

# With vision (multimodal)
./llama-mtmd-cli \
  -m gemma-4-e2b-it.Q4_K_M.gguf \
  --mmproj gemma-4-e2b-it.BF16-mmproj.gguf
```

## Training Details

### LoRA Configuration

```json
{
  "r": 64,
  "lora_alpha": 64,
  "target_modules": ["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"],
  "lora_dropout": 0,
  "task_type": "CAUSAL_LM"
}
```

### Training Curve

| Step | Loss |
|------|------|
| 10 | 8.147 |
| 20 | 1.074 |
| 30 | 0.469 |
| 100 | 0.099 |
| 500 | 0.043 |
| 1000 | 0.038 |
| 1875 | 0.040 |

### Dataset

2,000 synthetic examples covering:
- System tool invocations (file management, process control)
- Contact and calendar queries
- Device information retrieval
- Multi-step task planning
- Safety-aware refusals

## Evaluation

| Metric | Score |
|--------|-------|
| Tool Use Accuracy | **7/7 (100%)** |
| Correct Tool Selection | **7/7** |
| Parameter Extraction | **7/7** |

## Part of SnowClaw

SnowClaw is a privacy-first AI agent built for the [Google Gemma Hackathon](https://ai.google.dev/competition). It features:

- **On-device inference** via bundled Ollama
- **Dual security modes**: Paranoid (fully offline) / Smart Search (local + anonymous SearXNG)
- **E2E encrypted** communication between desktop and mobile
- **Hardware-aware** model selection (auto-detects CPU/GPU/RAM)

## Author

**Kennt Kim** — [Calida Lab](https://calidalab.ai)

## License

Apache 2.0 (following Gemma's license terms)