SalahALHaismawi/falcon-h1r-uae-laws

Falcon-H1R-7B UAE Laws LoRA Adapter

A LoRA adapter for tiiuae/Falcon-H1R-7B finetuned on UAE Federal Laws to provide legal analysis in IRAC format.

What is IRAC?

IRAC is a legal analysis framework:

• Issue: Identifies the legal question
• Rule: Cites the relevant law/article
• Application: Explains how the rule applies to the situation
• Conclusion: Provides the final answer

---

Quick Start Options

Method	Best For
GGUF + Ollama	Local deployment, easy setup
Python + LoRA	Custom integration, fine-tuning

---

GGUF + Ollama (Recommended)

The easiest way to run this model locally is with Ollama.

1. Install Ollama

curl -fsSL https://ollama.com/install.sh | sh

2. Download the GGUF

# Download from HuggingFace
wget https://huggingface.co/SalahALHaismawi/falcon-h1r-uae-laws/resolve/main/falcon-h1r-uae-laws.gguf

3. Create a Modelfile

Create a file named Modelfile:

FROM ./falcon-h1r-uae-laws.gguf

TEMPLATE """### Instruction:
{{ .Prompt }}

### Response:
"""

PARAMETER stop "### Instruction:"
PARAMETER temperature 0.1

4. Create and Run

# Create the model in Ollama
ollama create falcon-uae-laws -f Modelfile

# Run inference
ollama run falcon-uae-laws "What are the penalties for drunk driving in the UAE?"

Example Ollama Output

Issue: Drunk driving penalties. Rule: Under UAE Federal Decree-Law No. 14 of 2020
(amending the Penal Code) and the Federal Traffic Law, driving under the influence
of alcohol is a criminal offence punishable by imprisonment and or a fine...

---

Training Dataset

UAE Laws Q&A Dataset (IRAC Format)

• 9,477 samples of UAE legal Q&A pairs
• 100% IRAC format consistency
• 83.8% cite specific article numbers
• Human-written seed data (~1,000 pairs) + synthetic generation with human validation

from datasets import load_dataset
dataset = load_dataset("SalahALHaismawi/uae-laws-irac")

---

Example Outputs

Example 1: Criminal Law

Question: What are the penalties for drunk driving in the UAE?

Response:

Issue: Drunk driving penalties. Rule: Under UAE Federal Decree-Law No. 14 of 2020 (amending the Penal Code) and the Federal Traffic Law, driving under the influence of alcohol is a criminal offence punishable by imprisonment and or a fine, with penalties increasing if damage, injury, or death is caused. Application: Anyone who drives a vehicle while intoxicated may be punished by imprisonment and a fine as determined by the court. If the offence results in damage to property, bodily injury, or death, stricter penalties including longer imprisonment and higher fines may be imposed. Conclusion: Drunk driving in the UAE is a serious offence with penalties that escalate depending on the harm caused.

---

Example 2: Labor Law

Question: What happens if an employer terminates an employee without notice in the UAE?

Response:

Issue: Employer terminates without notice. Rule: Article 44 of Federal Decree-Law No. 33 of 2021 permits termination without notice only in specific circumstances. Article 47 provides compensation where termination is unlawful. Application: If an employer terminates an employee without notice and without a valid legal reason, the employee is entitled to compensation of up to three months' salary, in addition to notice pay and other contractual entitlements. Conclusion: An employer who unlawfully terminates an employee without notice must compensate the employee in accordance with UAE Labour Law.

---

Example 3: Property Law

Question: What are the rules regarding property ownership for foreigners in the UAE?

Response:

Issue: Can foreigners own property in the UAE? Rule: Foreigners may own property only in areas designated by the relevant Emirate, subject to applicable laws and regulations. Application: Emirate-level legislation determines the specific freehold or usufruct areas where foreign ownership is permitted. Conclusion: Foreigners may own property in the UAE only in designated areas and in accordance with Emirate-specific regulations.

---

Python Usage

from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
from peft import PeftModel
import torch

# Load with 8-bit quantization (required)
tokenizer = AutoTokenizer.from_pretrained("tiiuae/Falcon-H1R-7B", trust_remote_code=True)
bnb_config = BitsAndBytesConfig(load_in_8bit=True)

base_model = AutoModelForCausalLM.from_pretrained(
    "tiiuae/Falcon-H1R-7B",
    quantization_config=bnb_config,
    device_map="auto",
    max_memory={0: "15GiB"},  # Prevents CPU offloading issues
    trust_remote_code=True
)

# Load the LoRA adapter
model = PeftModel.from_pretrained(base_model, "SalahALHaismawi/falcon-h1r-uae-laws-lora")
model.eval()

# Generate response
question = "What are the penalties for drunk driving in the UAE?"
prompt = f"### Instruction:\n{question}\n\n### Response:\n"

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
with torch.no_grad():
    output = model.generate(
        **inputs,
        max_new_tokens=400,
        do_sample=False,
        pad_token_id=tokenizer.eos_token_id
    )

response = tokenizer.decode(output[0], skip_special_tokens=True)
print(response.split("### Response:")[-1].strip())

---

Training Details

Parameter	Value
Base Model	tiiuae/Falcon-H1R-7B
Method	LoRA (Low-Rank Adaptation)
Rank (r)	32
Alpha	64
Dropout	0.05
Target Modules	q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
Training Data	UAE Federal Laws

---

Available Files

File	Size	Description
falcon-h1r-uae-laws.gguf	~15GB	Quantized model for Ollama/llama.cpp
adapter_model.safetensors	~134MB	LoRA adapter weights
adapter_config.json	-	LoRA configuration

---

Important Notes

8-bit quantization is required for Python usage - The model produces incorrect outputs without it

GGUF works out of the box - No quantization setup needed when using Ollama

RAG Implementation is recommended: This fine-tune teaches the model IRAC format aligned with UAE legal text. For the most accurate legal output, implement a document retriever or web-crawler to reduce hallucinations and ground the truth.

Use the exact prompt format shown above for best results

Legal Disclaimer: This model is for educational purposes only. Always consult a qualified legal professional for actual legal advice.

---

Model Architecture

Falcon-H1R-7B is a hybrid architecture combining:

• Mamba (State Space Model) layers for efficient sequence processing
• Attention layers for complex reasoning

This makes it particularly efficient for long legal documents.

---

License

Apache 2.0 (following the base model license)

Author

Salah AlHaismawi