nadvkjdv/auralis-ap-urban-intelligence

Auralis AP Urban Intelligence (LoRA)

LoRA adapter on Qwen/Qwen2.5-1.5B-Instruct, used as the conversational layer of Auralis, an evidence-grounded operations platform for city infrastructure across Andhra Pradesh, India.

What this model does — and does not do

This adapter handles conversation only: greetings, follow-ups, and explaining the platform.

It is deliberately not the source of any fact shown to an operator. In Auralis, every reading — weather, air quality, river discharge, traffic speed, incident counts, news — is fetched from a live source by a tool, and the tool result is what reaches the screen. The system prompt forbids the model from stating a city measurement, because a 1.5B model asked for a temperature will produce a plausible one.

That split is the point. Grounding is enforced in the data layer, not by asking a language model to behave.

Usage

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

BASE = "Qwen/Qwen2.5-1.5B-Instruct"
ADAPTER = "nadvkjdv/auralis-ap-urban-intelligence"

tok = AutoTokenizer.from_pretrained(BASE)
model = AutoModelForCausalLM.from_pretrained(BASE, device_map="auto")
model = PeftModel.from_pretrained(model, ADAPTER)

messages = [
    {"role": "system", "content": "You are Auralis AI, the assistant inside a civic intelligence platform."},
    {"role": "user", "content": "What does the Trace page do?"},
]
prompt = tok.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
out = model.generate(**tok(prompt, return_tensors="pt").to(model.device), max_new_tokens=320)
print(tok.decode(out[0], skip_special_tokens=True))

Training

Base	Qwen/Qwen2.5-1.5B-Instruct
Method	LoRA (PEFT)
Rank	16
Alpha	32
Target modules	q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
Task	CAUSAL_LM

Runtime

Runs on CPU. Expect roughly 10–30 s per reply for a few hundred tokens on a typical CPU; a GPU brings that under a second. The host application treats the model as optional — when it is unavailable, tool-grounded answers are served without it.

Limitations

• Small model. It is not a reasoning engine and should not be asked to be one.
• Scoped to Andhra Pradesh civic vocabulary; other domains fall back to base behaviour.
• Do not use its output as a source of fact about any real-world measurement.

Licence

Apache 2.0, inheriting the base model's licence.