We fine-tuned a Llama-3.2-3B on 330 manually annotated queries, each mapped to $L, I, T$.

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
from clear.prompt import generate_query_prompt

device = "cuda" if torch.cuda.is_available() else "cpu"
tokenizer = AutoTokenizer.from_pretrained(
            "oscarwu/Llama-3.2-3B-CLEAR",
            use_auth_token=token)
model = AutoModelForCausalLM.from_pretrained(
    "oscarwu/Llama-3.2-3B-CLEAR",
    use_auth_token=token)
model.to(device)

query = "In Oakford, WA, in the Serpentine-Jarrahdale LGA, water scarcity and extreme heat are major challenges. What programs are in place to promote water efficiency and manage climate impacts in our region?"
test_prompt = generate_query_prompt(query)

inputs = tokenizer(
      test_prompt,
      padding=True,
      truncation=True,
      max_length=1024,
      return_tensors="pt"
).to(device)

outputs = model.generate(**inputs, max_new_tokens = 128, use_cache = True)
test_result = tokenizer.batch_decode(outputs)

The action delivery module processed decomposed resident queries ($L,I,T$) to generate reports utilized two source data: $S$ (ABS census, Wikipedia) and $D$ for policy documents.

• The $S$ for local government can be found db_lga.csv, and suburb data in db_suburb.csv.

uv run prepare_policy.py
# or
python prepare_policy.py

Suburb Schema

• Generates comprehensive reports for community and policy analysis
• Provides interactive map visualizations
• Automatically generates feedback emails to LGAs

uv run main.py \
    --device cuda \
    --query "In Oakford, WA..." \
    --gpt_model gpt-4