---
license: cc-by-nc-sa-4.0
base_model: Qwen/Qwen2.5-0.5B-Instruct
tags:
  - task-automation
  - local-llm
  - gguf
  - cpu-inference
  - fine-tuned
language:
  - en
library_name: llama-cpp
pipeline_tag: text-generation
---

# ACE 0.5B - Task Automation Model

Fine-tuned Qwen 0.5B for local task automation. Detects task types and generates execution plans.

**Code:** [GitHub](https://github.com/ansh0x/ace)

## Model Description

ACE is a 0.5B parameter language model fine-tuned for task automation. It can:
- Classify tasks (atomic, repetitive, clarification needed)
- Generate CLI commands for file operations
- Create execution plans with hotkeys
- Handle repetitive bulk operations

All inference runs on CPU - no GPU required.

## Model Files

| File | Size | Quant | Use Case |
|------|------|-------|----------|
| `ace-bf16.gguf` | 940MB | BF16 | **Recommended** - A bit slower inference, but better quality |
| `ace-q4-k-m.gguf` | 385MB | Q4_K_M | Faster inference |

## Training Details

**Base Model:** Qwen/Qwen2-0.5B  
**Method:** LoRA fine-tuning (r=16, alpha=32)  
**Dataset:** ~1000 custom task examples  
**Training:** 500-700 steps, learning_rate=3e-5  
**Quantization:** GGUF Q4_K_M with imatrix  

**Task Types:**
- Atomic tasks (single operations)
- Repetitive tasks (bulk processing)
- Clarification requests (ambiguous inputs)

**Data Format:**
```json
Input: {"task": "...", "directory": [...], "available_hotkeys": [...]}
Output: {"task_type": "atomic", "output": {"execution_plan": {...}}}
```

## Usage
- Right now the model is a bit unstable and intended for only experimental usages.
- Refer to the GitHub repo for installation and usage.

## Limitations

- Requires explicit file paths (no smart file search)
- Optimized for Linux commands (Should be able to work on Windows)
- CPU inference only (3-10 seconds on i3/i5)
- No visual understanding (text-only)
- English language only

## Performance

**Hardware benchmarks:**
- Intel i5 (2018+): 3-5 seconds per task
- Intel i3 (2015+): 5-10 seconds per task
- Older hardware: 30-90 seconds per task

## Bias and Ethics

**Known biases:**
- Training data focused on common developer workflows
- Linux command bias (more Linux than Windows examples)
- English-only (no multilingual support)

**Ethical considerations:**
- Model can generate destructive commands (file deletion)
- Users should review plans before execution
- No built-in safety checks for harmful operations

## License

**CC BY-NC-SA 4.0** (Non-commercial)

- ✅ Free for personal/research use
- ❌ Commercial use requires separate license
- ✅ Must provide attribution
- ✅ Derivatives must use same license

**Additional Restriction:**
Training of AI/ML models using these weights is prohibited without explicit written permission.

## Contact

- **Issues:** [GitHub Issues](https://github.com/ansh0x/ace/issues)
- **Discussions:** [GitHub Discussions](https://github.com/ansh0x/ace/discussions)

---

**More info:** [GitHub Repository](https://github.com/ansh0x/ace)
```