warehouseflow-gemma3-1b-it-gguf

WarehouseFlow Gemma 3 1B IT — GGUF

Model Gemma 3 1B IT yang telah di-finetune khusus untuk warehouse management tool calling, diekspor ke format GGUF untuk digunakan dengan llama.cpp, Ollama, LM Studio, Jan, dan runtime GGUF lainnya.

🎯 Use Case

• Warehouse Management System (WMS): Stok, transfer, PO, kapasitas
• Logistics AI Agent: Multi-step reasoning dengan 12 tools
• Edge Deployment: Berjalan di laptop/RAM terbatas (Q4_K_M ~0.75 GB)

📊 Varian Kuantisasi

🛠️ Tools (12 Function Calls)

Model ini mengenali dan memanggil 12 tools berikut:

1. get_stock(sku, warehouse_id) - Cek stok SKU di warehouse
2. transfer_stock(sku, quantity, source_warehouse_id, destination_warehouse_id) - Transfer stok
3. list_warehouses(region=None) - List semua warehouse
4. get_stock_history(sku, warehouse_id, days=30) - Riwayat stok
5. create_purchase_order(sku, quantity, supplier_id, warehouse_id) - Buat PO
6. get_supplier_info(supplier_id) - Info supplier
7. check_low_stock_alerts(warehouse_id) - Alert stok rendah
8. get_warehouse_capacity(warehouse_id) - Kapasitas warehouse
9. search_sku_by_name(query) - Cari SKU by nama
10. get_shipping_status(order_id) - Status pengiriman
11. estimate_delivery_time(source_warehouse_id, destination) - ETA pengiriman
12. get_supplier_lead_time(supplier_id) - Lead time supplier

💬 Format Output

<think>
[Reasoning block - optional untuk query kompleks]
</think>
<tool_call>
{"name": "get_stock", "arguments": {"sku": "ZX-1042", "warehouse_id": "JKT-01"}}
</tool_call>

🚀 Penggunaan

1. llama.cpp (CLI)

./llama-cli -m warehouseflow-gemma3-1b-gguf/gemma-3-1b-it.Q5_K_M.gguf \
            -p "Cek stok ZX-1042 di JKT-01" \
            -n 256 --temp 0

2. Ollama

Buat Modelfile:

FROM hf.co/Ugisr/warehouseflow-gemma3-1b-it-gguf:F16
PARAMETER temperature 0 PARAMETER num_ctx 2048 PARAMETER stop "<end_of_turn>"
TEMPLATE """<start_of_turn>user {{ .System }}
{{ .Prompt }}<end_of_turn> <start_of_turn>model """
SYSTEM """You are a warehouse management AI assistant with access to these tools:
    1. get_stock(sku: str, warehouse_id: str)
    2. transfer_stock(sku: str, quantity: int, source_warehouse_id: str, destination_warehouse_id: str)
    3. list_warehouses(region: str = None)
    4. get_stock_history(sku: str, warehouse_id: str, days: int = 30)
    5. create_purchase_order(sku: str, quantity: int, supplier_id: str, warehouse_id: str)
    6. get_supplier_info(supplier_id: str)
    7. check_low_stock_alerts(warehouse_id: str)
    8. get_warehouse_capacity(warehouse_id: str)
    9. search_sku_by_name(query: str)
    10. get_shipping_status(order_id: str)
    11. estimate_delivery_time(source_warehouse_id: str, destination: str)
    12. get_supplier_lead_time(supplier_id: str)
Reasoning & output rules:
    • For a simple single lookup, answer with exactly one tool call and no reasoning block.
    • For planning, several independent lookups, or conditional decisions, FIRST write reasoning inside ..., then act.
    • You may emit one or more <tool_call>...</tool_call> blocks. Emit several ONLY when calls are independent and every argument is already known.
    • Each <tool_call> must contain exactly one JSON object: {"name": "", "arguments": }.
    • If a decision depends on a value you do not have yet, call the tool that fetches it FIRST; never guess.
    • If the request is ambiguous, lacks required info, or is outside warehouse operations, do NOT emit any tool call; ask a short clarification in plain text.
    • Only use the tools listed above. Never invent a tool name or an argument value.
    • Respond in the same language as the user's query."""

Lalu:

ollama create warehouseflow -f Modelfile
ollama run warehouseflow "Cek stok ZX-1042 di JKT-01."

3. Python (llama-cpp-python)

from llama_cpp import Llama

llm = Llama(
    model_path="gemma-3-1b-it.Q5_K_M.gguf",
    n_ctx=2048,
    n_gpu_layers=-1,  # gunakan GPU jika ada
)

output = llm(
    "Cek stok ZX-1042 di JKT-01.",
    max_tokens=256,
    temperature=0,
)
print(output["choices"][0]["text"])

4. LM Studio / Jan

• Copy file .gguf ke folder models
• Load dari UI
• Set temperature = 0 untuk deterministic output

📈 Benchmark (LogiBench)

📝 Catatan

• Model ini bukan general-purpose LLM — dioptimalkan untuk domain warehouse/logistics
• Gunakan temperature = 0 untuk output deterministic (penting untuk tool calling)
• Untuk query kompleks, model akan mengeluarkan blok <think> sebelum memanggil tool
• Model akan MENOLAK pertanyaan di luar domain logistik (cuaca, politik, dll)

⚠️ Disclaimer

Model ini adalah fine-tuned Gemma 3 1B IT. Silakan ikuti Gemma Terms of Use dari Google.