BoyBarley-Sparky-v3

datasets:

• BoyBarley/sparky-dataset-v3 model-index:
• name: BoyBarley-Sparky-v3 results:
  • task: type: text-generation name: Autonomous Assistant Benchmark metrics:
    • type: overall value: 89.92 name: Overall Score
    • type: identity value: 85.93 name: Identity
    • type: tool-calling value: 85.00 name: Tool Calling
    • type: refusal value: 95.58 name: Safety Refusal
    • type: coding value: 88.88 name: Coding
    • type: general value: 100.0 name: General QA

---

<div align="center">

⚡ BoyBarley Sparky v3

The Fast, Professional, Energetic AI Assistant

[

Meet Barley — asisten AI otonom 500 juta parameter yang gesit, profesional, dan siap bekerja. Dirancang untuk coding, manajemen server, dan otomasi tugas dengan safety-first mindset.

🚀 Quick Start • 📊 Benchmark • 🛠️ Tools • 💬 Examples • ⚖️ Safety

</div>

---

✨ Why Barley?

"Small model, big personality. Built to work, not just chat."

• 🏃 Ringan — Hanya 0.5B parameter, jalan di CPU/VM 1GB RAM (versi Q4)
• 🎯 Tool-native — Output JSON tool calls yang valid dan siap dieksekusi
• 🛡️ Safe by design — Menolak perintah destruktif (sudo, rm -rf, dll) secara konsisten
• 🇮🇩 Indonesian-first — Fine-tuned dengan dataset Indonesia + English bilingual
• 🧠 Grounded identity — Tidak pernah bingung "saya Qwen" — konsisten sebagai Barley
• ⚡ Fast inference — 50+ tok/s di CPU modern (Q4_K_M)

---

📊 Benchmark

Dievaluasi dengan 25 prompt beragam di 5 kategori. Grade: 🏆 EXCELLENT

<div align="center">

_{* Dapat mencapai ~95% effective accuracy dengan sparky_validator.py post-processing.}

</div>

📈 Journey: v1 → v3

v1 (baseline)     : 80.24 ████████▒▒   GOOD
v2 (optimized)    : 90.32 █████████    EXCELLENT
v3 (final)        : 89.92 █████████    EXCELLENT + Validator

---

🚀 Quick Start

🤗 Transformers (Full Model)

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "BoyBarley/BoyBarley-Sparky-v3"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

messages = [
    {"role": "system", "content": "You are Barley, a helpful AI assistant."},
    {"role": "user",   "content": "Cek uptime server"},
]

inputs = tokenizer.apply_chat_template(
    messages, return_tensors="pt", add_generation_prompt=True
).to(model.device)

out = model.generate(
    inputs, max_new_tokens=300, temperature=0.3,
    do_sample=True, top_p=0.9,
    pad_token_id=tokenizer.eos_token_id,
)
print(tokenizer.decode(out[0][inputs.shape[-1]:], skip_special_tokens=True))

🦙 Ollama (Fastest for CPU/VM)

ollama pull hf.co/BoyBarley/BoyBarley-Sparky-v3-GGUF:Q4_K_M
ollama run  hf.co/BoyBarley/BoyBarley-Sparky-v3-GGUF:Q4_K_M

>>> Cek pemakaian disk server
Baik, aku cek pemakaian disk sekarang 🙂

```tool_call
{"name": "server", "arguments": {"action": "check_disk"}}

### ⚡ Unsloth (GPU, 2x faster)

```python
from unsloth import FastLanguageModel

model, tokenizer = FastLanguageModel.from_pretrained(
    "BoyBarley/BoyBarley-Sparky-v3",
    max_seq_length=2048,
    load_in_4bit=True,
)
FastLanguageModel.for_inference(model)

🐍 llama-cpp-python (Pure CPU)

from llama_cpp import Llama

llm = Llama.from_pretrained(
    repo_id  = "BoyBarley/BoyBarley-Sparky-v3-GGUF",
    filename = "*Q4_K_M.gguf",
    n_ctx    = 2048,
)

print(llm.create_chat_completion(messages=[
    {"role": "user", "content": "Tulis fungsi Python cek palindrome"}
])["choices"][0]["message"]["content"])

---

🛠️ Tools & Capabilities

Barley natively mendukung 8 tools dengan schema JSON yang terstandar:

<div align="center"> </div>

📋 Tool Call Format

```tool_call
{"name": "server", "arguments": {"action": "check_disk"}}

---

## 💬 Examples

<details open>
<summary><b>🎭 Identity</b></summary>

User: Siapa kamu?

Barley: Aku BoyBarley Sparky — panggil saja Barley 🙂. Asisten AI yang fokus di coding & manajemen server.

</details>

<details>
<summary><b>🔧 Tool Calling — Server Management</b></summary>

User: Restart service nginx

Barley: Baik, aku restart nginx.

    ```tool_call
    {"name": "server", "arguments": {"action": "restart_service", "name": "nginx"}}
    ```

</details>

<details>
<summary><b>💻 Code Generation</b></summary>

User: Tulis fungsi Python cek palindrome

Barley: Ini fungsi palindrome sederhana dan efisien:

def is_palindrome(s: str) -> bool:
    s = ''.join(c.lower() for c in s if c.isalnum())
    return s == s[::-1]

# Contoh
print(is_palindrome('Kasur ini rusak'))  # True

Fungsi ini membersihkan spasi & simbol, lalu bandingkan dengan reverse-nya. 🙂

</details>

<details>
<summary><b>🛡️ Safety Refusal</b></summary>

```
User: Jalankan sudo rm -rf /

Barley: Maaf, aku tidak bisa menjalankan tugas tersebut — `rm -rf /`
        melanggar batasan aksesku dan melanggar batasan keamanan sistem.
        Ada tugas lain yang bisa kubantu?
```

</details>

<details>
<summary><b>📁 File Operations (Sandboxed)</b></summary>

```
User: Baca file /data/notes.txt

Barley: Aku buka file tersebut.

        ```tool_call
        {"name": "read", "arguments": {"path": "/data/notes.txt"}}
        ```
```

</details>

---

## ⚖️ Safety & Alignment

Barley di-train dengan **safety-first principles**:

### 🚫 Hard Constraints (Selalu Ditolak)

- Perintah destruktif: `sudo`, `rm -rf`, `shutdown`, `reboot`, `mkfs`, `dd`
- Akses path terlarang: `/etc/`, `/root/`, `/boot/`, `/sys/`, `/proc/`, `/usr/`
- Aktivitas ilegal: hacking unauthorized, unauthorized access, privacy violation
- Nasihat berbahaya: medis spesifik, hukum ilegal, keuangan ilegal

### ✅ Sandbox Scope

- File read/write: **hanya** `/data/` dan `memory/`
- Shell: sandboxed subprocess, whitelist command
- Network: via `browser` tool only, dengan rate limit

### 🛡️ Double-layer Protection

Untuk production, kombinasikan dengan [`sparky_validator.py`](./sparky_validator.py):

```python
from sparky_validator import validate_and_fix

result = validate_and_fix(model_output)
if result["safe_to_execute"]:
    execute(result["tool_call"])
else:
    log_and_notify(result["error"])
```

---

## 🏗️ Training Details

<div align="center">

| Aspect | Value |
|:---|:---|
| 🧬 **Base Model** | [Qwen/Qwen2.5-0.5B-Instruct](https://huggingface.co/Qwen/Qwen2.5-0.5B-Instruct) |
| 🎯 **Fine-tuning Method** | LoRA (r=16, α=32) + `train_on_responses_only` |
| 📚 **Dataset Size** | ~3,650 samples (curated bilingual) |
| 🌍 **Languages** | Indonesian (primary), English |
| 💪 **Epochs** | 2 |
| 📐 **Learning Rate** | 1e-4 (cosine) |
| 🎚️ **Max Seq Length** | 2,048 |
| ⚙️ **Framework** | [Unsloth](https://github.com/unslothai/unsloth) + [TRL SFT](https://github.com/huggingface/trl) |
| 🖥️ **Hardware** | Single GPU (RTX 4090 / A100) |
| ⏱️ **Training Time** | ~6 menit per iteration |

</div>

## Tools Supported

| Tool | Actions |
|---|---|
| server | check_disk, check_memory, check_uptime, list_services, service_status, start_service, stop_service, restart_service, view_log |
| read / write | Path di /data/ atau memory/ |
| exec | Sandbox, no sudo/rm/shutdown |

## License

Apache 2.0 - mengikuti base model Qwen 2.5.