Qwen 2.5 0.5B Tool-Augmented Traffic SOP Dispatcher (HamzaBoy/qwen2.5-0.5b-traffic-sop) is a specialized 500M parameter LLM fine-tuned via SFT + QLoRA to act as Stage 4: Autonomous Traffic Officer Dispatcher for Smart City Command Centers.

It evaluates real-time traffic telemetry, invokes Agentic Tools (such as Dijkstra Shortest Path Routing over city road graphs), and selects optimal Standard Operating Procedure (SOP) macro-actions (VERIFY, DISPATCH, RESOLVE, REJECT, ESCALATE).

🛠️ Model Capabilities & Agentic Tools

The model is trained to execute multi-step tool calls before issuing a final structured JSON decision:

Tool Name	Operation & Purpose
`calculate_shortest_route`	Calculates shortest path ($\text{km}$) & ETA ($\text{mins}$) using Dijkstra's Algorithm over city road graph weighted by live congestion.
`query_available_units`	Queries available patrol bikes, interceptor vans, and heavy tow trucks near police station jurisdiction.
`check_junction_cctv`	Fetches live visual camera analytics (lane blockage, stalled vehicles, visibility %).
`issue_signal_override`	Triggers automated Green Corridor traffic light priority for emergency clearance.
`broadcast_traffic_advisory`	Publishes diversion alerts to public VMS boards and navigation systems.

📊 Training Specifications

Base Model: Qwen/Qwen2.5-0.5B-Instruct
Developer: Ameer Hamza Khan (@HamzaBoy)
Dataset: 5,000 tool-augmented real-world traffic trajectories incorporating weather (HEAVY_RAIN, WATERLOGGING), speed drop %, ambulance flags, and citizen reliability scores.
Hardware: Google Colab T4 GPU (4-bit QLoRA, r=16, alpha=32).
Convergence: Training loss dropped from 1.3902 down to 0.0851 (~99% accuracy on SOP rules & JSON tool syntax).

💻 Usage Example

python

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

BASE_MODEL = "Qwen/Qwen2.5-0.5B-Instruct"
ADAPTER_REPO = "HamzaBoy/qwen2.5-0.5b-traffic-sop"

tokenizer = AutoTokenizer.from_pretrained(ADAPTER_REPO)
base_model = AutoModelForCausalLM.from_pretrained(BASE_MODEL, torch_dtype=torch.float16, device_map="auto")
model = PeftModel.from_pretrained(base_model, ADAPTER_REPO)

system_prompt = "You are an AI Traffic Officer Dispatcher. Available tools: [calculate_shortest_route, query_available_units, check_junction_cctv, issue_signal_override, broadcast_traffic_advisory]. Select optimal action: VERIFY, DISPATCH, RESOLVE, REJECT, ESCALATE."

prompt = "Incident Alert TICK-BLR-0941: Station=Bellandur, Junction=Silk Board Flyover, Weather=HEAVY_RAIN, SpeedDrop=88%, AmbulanceBlocked=TRUE."

messages = [
    {"role": "system", "content": system_prompt},
    {"role": "user", "content": prompt}
]

inputs = tokenizer(tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True), return_tensors="pt").to("cuda")
outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

🔄 Human-in-the-Loop (HITL) & Continuous Alignment

The model operates with a Softmax Confidence Gate:

If $\text{Softmax Confidence} \ge 0.80$ AND action $\neq \text{ESCALATE} \implies$ AUTONOMOUS EXECUTION.
If $\text{Softmax Confidence} < 0.80$ OR action $= \text{ESCALATE} \implies$ TRIGGERS HITL OFFICER MODAL.

Officer approvals and manual overrides are automatically logged as DPO Preference Pairs for continuous background micro-tuning.