uae-vat-llama3.2-3b

UAE VAT Assistant — Llama-3.2-3B (fine-tuned)

A small language model fine-tuned to answer UAE Value Added Tax (VAT) questions, based on Federal Tax Authority (FTA) rules. Built by fine-tuning meta-llama/Llama-3.2-3B-Instruct with QLoRA using LLaMA-Factory, then merging the adapter.

⚠️ Educational / demo only — not tax advice. Small fine-tuned models can be inaccurate and may state outdated or incorrect figures. Verify everything against the FTA (tax.gov.ae) or a registered UAE tax agent.

Intended use

• Answering general questions about UAE VAT: the 5% standard rate, registration thresholds, zero-rated vs exempt supplies, filing, and input-tax recovery.
• A learning / portfolio demonstration of domain fine-tuning.

Out of scope: authoritative tax rulings, filings, multi-step numeric tax computations, or any decision made without professional verification.

How to use

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "adarshsm28/uae-vat-llama3.2-3b"
tok = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.bfloat16, device_map="auto")

msgs = [{"role": "user", "content": "When must a business register for VAT in the UAE?"}]
inputs = tok.apply_chat_template(msgs, add_generation_prompt=True, return_tensors="pt").to(model.device)
out = model.generate(inputs, max_new_tokens=256, temperature=0.3, top_p=0.9, repetition_penalty=1.1)
print(tok.decode(out[0][inputs.shape[-1]:], skip_special_tokens=True))

Training

• Base: meta-llama/Llama-3.2-3B-Instruct (~3.2B params)
• Method: QLoRA (4-bit) via LLaMA-Factory, adapter then merged
• Data: Alpaca-format UAE VAT instruction/response pairs, grounded in FTA rules (registration thresholds, supply classification, filing, invoicing, 2026 updates)
• Precision: bf16

Key facts the model targets

• Standard VAT rate: 5%
• Registration thresholds: AED 375,000 (mandatory), AED 187,500 (voluntary) — these are the only thresholds
• Zero-rated (0%): exports, international transport, first supply of new residential property, specified healthcare and education
• Exempt: certain financial services, subsequent residential property, bare land, local passenger transport

Limitations & known issues

• Hallucination: with limited training data the model can invent figures or mislabel supplies (e.g. treating everyday goods as zero-rated). Do not rely on it.
• Staleness: UAE VAT rules change (e.g. 2026 law amendment, penalty reform, e-invoicing rollout). Re-verify against current FTA guidance.
• Not grounded in retrieval — for a reliable tool, pair it with RAG over the FTA guides so answers cite source text.

License

Governed by the Llama 3.2 Community License.