⚡ ZOZ-Function-Master-3B

Autonomous Function Calling & Agentic Tool Execution Engine for Edge Hardware

---

📌 Overview

ZOZ-Function-Master-3B is an instruction-tuned, LoRA-merged 3-billion-parameter language model purpose-built for high-reliability Function Calling, Agentic Tool Invocation, and Strict JSON Output Generation.

Trained to eliminate schema hallucinations, key drifting, and syntax invalidations, this model achieves parity with leading 7B architectures while maintaining the execution speed and low-latency footprint required for local edge deployment, embedded agents, and multi-tenant production backends.

---

📊 Empirical Benchmarks

All models were evaluated under an identical automated evaluation scaffold using deterministic decoding parameters (temperature=0.0, do_sample=False, max_new_tokens=256). Responses are parsed programmatically via an Abstract Syntax Tree (AST) validator with zero tolerance for schema malformations, invalid primitive types, or unrequested tool invocations.

Model Name	Parameters	Class	AST Strictness	Overall Pass Rate (%)
🚀 ZOZ-Function-Master-3B (Ours)	3B	Edge	Strict AST / JSON	83.3%
Qwen2.5-7B-Instruct	7B	Mid-Scale	Strict AST / JSON	83.3%
Mistral-7B-Instruct-v0.3	7B	Mid-Scale	Strict AST / JSON	83.3%
Qwen2.5-3B-Instruct	3B	Edge	Strict AST / JSON	83.3%
Gemma-2-2b-it	2.6B	Edge	Partial / Drift	66.7%
Phi-3.5-mini-instruct	3.8B	Edge	Format Drift	33.3%

🔬 Granular Task Breakdown

Evaluation Scenario	Test Criteria	ZOZ-3B (Ours)	Phi-3.5-mini (3.8B)	Gemma-2-2b-it
Single Direct Call	Key extraction & parameter mapping	✅ PASS	❌ FAIL	✅ PASS
Strict Numeric Types	Integer bounds & float scalar precision	✅ PASS	✅ PASS	✅ PASS
Parallel Execution	Concurrent multi-tool requests	✅ PASS	❌ FAIL	❌ FAIL
Nested Schema	Deep JSON objects & array-of-objects	✅ PASS	❌ FAIL	✅ PASS
Negative Trigger (No-Op)	Rejecting irrelevant tools on factual queries	✅ PASS	✅ PASS	✅ PASS
Injection Defense	Hardening against embedded malicious calls	❌ FAIL	❌ FAIL	❌ FAIL

---

🎯 Key Architectural Strengths

• Zero-Drift Type Enforcement: Eliminates common small-model defects where numbers or booleans are wrapped as strings (e.g., outputs {"count": 3, "active": false} instead of {"count": "3", "active": "false"}).
• Concurrent Parallelism: Accurately emits multiple <tool_call> blocks in a single inference step when queries require simultaneous data fetching.
• Deeply Nested JSON: Formats complex hierarchical structures, including lists of dictionaries, without truncating closing brackets.
• Low Hardware Footprint: Consumes less than 2.2 GB VRAM under 4-bit quantization (GGUF / AWQ), enabling local execution on single GPUs, Apple Silicon, and edge microservers.

---

🚀 Quickstart & Inference

1. Using Transformers

import json
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "z51722369/ZOZ-Function-Master-3B"

tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16 if torch.cuda.is_bf16_supported() else torch.float16,
    device_map="auto"
)

# Define your tool schema
tools = [
    {
        "type": "function",
        "function": {
            "name": "book_flight",
            "description": "Reserve passenger airline seats.",
            "parameters": {
                "type": "object",
                "properties": {
                    "flight_number": {"type": "string"},
                    "seats": {"type": "integer"},
                    "max_price": {"type": "number"}
                },
                "required": ["flight_number", "seats", "max_price"]
            }
        }
    }
]

messages = [
    {"role": "system", "content": "You are a precise tool execution agent. Return clean JSON calls inside <tool_call> tags when required."},
    {"role": "user", "content": "Reserve 3 seats for flight MS-777 with a price limit of $450.50."}
]

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

with torch.no_grad():
    outputs = model.generate(
        **inputs,
        max_new_tokens=256,
        temperature=0.01,
        do_sample=False
    )

print(tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True))