rajeshthangaraj1/uae_rule_book_QA_assistant

RegulaUAE-1.2B - UAE Rulebook Q&A Assistant - Finetuned LFM2 Model

Model ID: rajeshthangaraj1/uae_rule_book_QA_assistant Base Model: [unsloth/LFM2-1.2B](https://docs.unsloth.ai/) ---

📘 Model Overview

RegulaUAE-1.2B is a domain-specific conversational language model fine-tuned to answer questions strictly grounded in the UAE Central Bank Rulebook (Banking Regulations).

The model is designed to support regulatory, compliance, and educational use cases, with a strong focus on reduced hallucination within the UAE banking domain. Tested against CBUAE regulatory queries.

Coverage includes:

• —Capital Adequacy
• —Licensing & Authorization
• —Corporate Governance
• —Risk Management
• —Compliance & Supervisory Frameworks

🔍 Key Characteristics

• —Domain: UAE Central Bank – Banking Regulations
• —Dataset Size: 500,000+ question–answer pairs
• —Language: English (complete rulebook coverage)
• —Precision: bfloat16
• —Task Type: Domain-specific legal & regulatory Q&A
• —Framework: Hugging Face transformers
• —Pipeline: text-generation with chat template support

🎯 Intended Use Cases

Regulatory & Legal Q&A

• —What is the relationship between Decree Law No. (20) of 2018 and Cabinet Decision No. (10) of 2019?
• —What minimum capital ratios are specified under Article (2)?

Compliance & Risk Teams

• —Regulatory validation and internal compliance support

Education & Research

• —Learning UAE banking regulations in a conversational format

AI & FinTech Development

• —Base model for regulation-aware RAG systems

⚠️ Limitations

• —Hallucination Risk: Without retrieval-augmented generation (RAG), the model may generate plausible but incorrect answers in edge cases.
• —Domain Scope: Limited to UAE Central Bank banking regulations only.
• —Numerical Accuracy: Percentages, ratios, and article references should be verified against the official rulebook.

📊 Dataset Creation

Source Data

Publicly available content from the official UAE Central Bank Rulebook: https://rulebook.centralbank.ae

Preprocessing

• —Scraped and cleaned official rulebook content
• —Segmented into ~65,000+ semantically aligned text chunks
• —Average chunk size: ~500 characters
• —Preserved articles, clauses, and legal definitions

Q&A Generation

Each chunk was used as grounded context to generate question–answer pairs.

Dataset structure:

{
  "context": "Rulebook text chunk",
  "question": "Regulatory question",
  "answer": "Answer grounded in the context"
}

🧪 Example Usage (Transformers)

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model_id = "rajeshthangaraj1/uae_rule_book_QA_assistant"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    device_map="auto",
    torch_dtype=torch.bfloat16
)

messages = [
    {"role": "system", "content":
     "You are an assistant specialized in the UAE Central Bank Rulebook. "
     "Only answer based on the UAE Rulebook. "
     "If the answer is not in the Rulebook, reply 'Not found in UAE Rulebook'."},
    {"role": "user", "content":
     "According to the UAE Central Bank Rulebook – Capital Adequacy Section, "
     "what does Article (2) specify about minimum capital ratios?"}
]

inputs = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt=True,
    return_tensors="pt"
).to(model.device)
inputs.pop("token_type_ids", None)

outputs = model.generate(**inputs, max_new_tokens=128)
answer = tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True)
print(answer)

🖥️ Example Usage (Gradio)

import gradio as gr

def chat_with_model(message, history):
    # (your chat_with_model function here)
    ...

gr.ChatInterface(fn=chat_with_model, title="UAE Rulebook QA Assistant").launch()

🔧 Technical Details

• —Base Model: LFM2-1.2B
• —Fine-tuning: LoRA adapters
• —Precision: bfloat16
• —Training Stack: Hugging Face transformers + accelerate

🛣️ Roadmap

• —Retrieval-Augmented Generation (RAG) integration
• —Arabic language support
• —Enhanced hallucination reduction and safety controls
• —Productization for compliance-critical environments

✍️ Author: @rajeshthangaraj1 📅 Last Updated: 2026