$ ollama run ckacor/hawk-1.5b \
>"Get GPU temp and check system uptime"
Execution Ready
---
## Technical architecture & capabilities
**Hawk-1.5B** by **ckacor** is a specialized agent model built on `Qwen2.5-1.5B-Instruct` and fine-tuned using the **QLoRA** methodology.
Its purpose is to parse natural-language requests (EN/RU) and convert them into structured, executable **JSON function calls** for system administration tasks. It does not produce conversational text — it produces operational data.
### Out-of-the-box toolchain
| Tool function | Description | Typical use case |
|---|---|---|
| `get_gpu_status` | Returns GPU metrics: VRAM, load, temperature | Inference node monitoring |
| `execute_terminal_command` | Runs an arbitrary Bash command | DevOps automation, daemon management |
| `read_file` | Reads the contents of a specified file path | Config inspection, log parsing |
Example prompts it handles
| Prompt | Model output |
|---|---|
| 🇬🇧 "Check GPU utilization" | `{"tool": "get_gpu_status", "arguments": {}}` |
| 🇬🇧 "Uptime?" | `{"tool": "execute_terminal_command", "arguments": {"command": "uptime"}}` |
| 🇷🇺 "Проверь дисковое пространство" | `{"tool": "execute_terminal_command", "arguments": {"command": "df -h"}}` |
| 🇷🇺 "Покажи конфиг config.json" | `{"tool": "read_file", "arguments": {"filepath": "config.json"}}` |
> No formal latency/accuracy benchmarks have been measured yet — numbers will be added here once evaluated, rather than estimated.
---
## Instant deployment with Ollama
1. **Download the GGUF** from the **Files and versions** tab.
2. **Create a Modelfile:**
```dockerfile
FROM ./hawk-1.5b-q8_0.gguf
PARAMETER temperature 0.1
PARAMETER top_p 0.95
PARAMETER repeat_penalty 1.1
SYSTEM """
You are HAWK-1.5B,
a Linux automation AI agent developed by ckacor.
Your task is to transform user requests
into accurate function calls.
"""
```
3. **Build and run:**
```bash
ollama create hawk-1.5b -f Modelfile
ollama run hawk-1.5b
```
```
> Check GPU temperature
{
"tool": "get_gpu_status",
"arguments": {}
}
```
### Via Transformers
```python
from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained("ckacor/my-lora-model")
tokenizer = AutoTokenizer.from_pretrained("ckacor/my-lora-model")
```
---
## Files and versions
> Only `q8_0` is available right now. Additional quantizations will be added once generated.
| File | Description |
|---|---|
| `hawk-1.5b-q8_0.gguf` | Maximum precision (available) |
| `hawk-1.5b-q5_k_m.gguf` | Higher quality, smaller size — coming soon |
| `hawk-1.5b-q4_k_m.gguf` | Recommended for most users — coming soon |
---
## Recommended runtime settings
| Parameter | Value |
|---|---|
| Temperature | 0.1 – 0.3 |
| Top P | 0.9 – 0.95 |
| Context | depends on hardware |
| Quantization | Q4_K_M recommended once available |
| Mode | deterministic generation |
---
## Training details
| | |
|---|---|
| **Base model** | Qwen2.5-1.5B-Instruct |
| **Method** | QLoRA |
| **Data** | Custom dataset of tool-calling examples for system administration tasks |
| **Languages** | Russian, English |
---
## Roadmap
Model
Focus
Status
Hawk-1.5B
General tool-use / sysadmin
✅ released (this repo)
Hawk-1.5B-Instruct
Broader instruction following
🔜 planned
Hawk-1.5B-Reasoning
Multi-step reasoning
🔜 planned
Hawk-1.5B-Code
Code generation
🔜 planned
Hawk-1.5B-Vision
Image understanding
🔜 planned
---
## Limitations
- Not designed for long creative writing
- Not a replacement for large general-purpose LLMs
- Tool execution requires an external agent layer — this model only outputs the JSON, it does not execute commands itself
- Performance depends on fine-tuning data quality
- May occasionally switch languages mid-response
- No formal benchmark numbers have been published yet
---
## Safety notes
HAWK-1.5B only generates commands and actions — it does not execute them. The execution layer should always:
- validate commands before running them
- restrict permissions to the minimum necessary
- use sandboxing where possible
- require confirmation for destructive operations
---
## About ckacor
ckacor develops lightweight AI systems focused on:
- local artificial intelligence
- machine learning experiments
- Linux automation
- efficient models for limited hardware
```
ckacor AI
|
▼
HAWK Family
|
├── Hawk-1.5B
|
└── Future Models
```
---
## License
This model follows the license specified in the model repository (Apache 2.0). Please check the license before commercial deployment.
---
## Citation
```bibtex
@misc{ckacor2026hawk,
title = {Hawk-1.5B: a LoRA fine-tune of Qwen2.5-1.5B for sysadmin tool-use},
author = {ckacor},
year = {2026},
url = {https://huggingface.co/ckacor/my-lora-model}
}
```
---
**🦅 HAWK-1.5B**
Built by ckacor
*Local AI. Efficient automation.*
