A Persian domain-adapted language model for answering questions about the "Operational By-law and General Conditions of Water and Wastewater Tariffs".

Base model

Qwen/Qwen2.5-1.5B-Instruct

Training data

Records: 274
Training records: 246
Validation records: 28
Format: conversational prompt-completion JSONL
Split: group-aware by source section to reduce source leakage

Training method

Method: QLoRA / supervised fine-tuning
Quantization: 4-bit NF4 with double quantization
LoRA rank: 16
LoRA alpha: 32
LoRA dropout: 0.05
Target modules: all-linear
Learning rate: 1e-04
Requested epochs: 10
Maximum sequence length: 512
GPU: Tesla T4

Validation generation metrics

Base model

Exact match: 0.0000
Token F1: 0.2044
ROUGE-L F1: 0.1670

Fine-tuned model

Exact match: 0.0000
Token F1: 0.3623
ROUGE-L F1: 0.3196

The automatic metrics measure textual similarity and do not establish legal correctness. Manual review is required for authoritative use.

Usage

python

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "Hesam760/persian-water-regulations-qwen2.5-1.5b"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    device_map="auto",
    dtype=torch.float16,
)

messages = [
    {
        "role": "system",
        "content": (
            "شما دستیار تخصصی آیین‌نامه عملیاتی و شرایط عمومی تعرفه‌های "
            "آب و فاضلاب هستید. فقط بر اساس مفاد آیین‌نامه پاسخ دهید."
        ),
    },
    {
        "role": "user",
        "content": "هدف آیین‌نامه چیست؟",
    },
]

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

output_ids = model.generate(
    **inputs,
    max_new_tokens=160,
    do_sample=False,
)

new_tokens = output_ids[0, inputs["input_ids"].shape[-1]:]
print(tokenizer.decode(new_tokens, skip_special_tokens=True))

Intended use

Demonstration of Persian domain adaptation
Regulatory question answering
Educational and research use

Limitations

The model is trained on a small, single-document dataset.
It can reproduce outdated, incomplete, or incorrectly extracted provisions.
It may hallucinate provisions or section numbers.
It is not a legal authority and must not be used as a substitute for the official regulation or professional advice.
Verify the redistribution rights of the source-derived training dataset before publishing the dataset itself.