HamzaBoy

qwen2.5-0.5b-traffic-sop

README

License: apache-2.0

🛠️ Model Capabilities & Agentic Tools

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

Table with columns: Tool Name, Operation & Purpose
Tool Name	Operation & Purpose
`calculate_shortest_route`	Calculates shortest path ( $km$ ) & ETA ( $mins$ ) 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 $Softmax Confidence \geq 0.80$ AND action $\neq = ESCALATE ⟹$ AUTONOMOUS EXECUTION.
If $Softmax Confidence < 0.80$ OR action $= ESCALATE ⟹$ TRIGGERS HITL OFFICER MODAL.

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

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Container

Run this model inference with full control and performance in your environment.

Learn more

Model Details

Model Provider

HamzaBoy

Model Tree

Base

Qwen/Qwen2.5-0.5B-Instruct

Adapter

this model

Input Modalities

Text

Output Modalities