peterrs23/audit-v-qwen-3b

Audit V — Qwen2.5 3B Instruct QLoRA Adapter

Audit V is an AI-powered financial, taxation, and audit intelligence platform. This repository contains the official 4-bit QLoRA PEFT adapter fine-tuned on the Audit V Indian Financial & Tax Instruction Dataset.

• Repository: peterrs23/audit-v-qwen-3b
• Base Model: Qwen/Qwen2.5-3B-Instruct
• Developer: Audit V AI Engineering Team

---

📌 Model Overview & Purpose

This LoRA adapter enhances Qwen/Qwen2.5-3B-Instruct with domain-specific knowledge of:

1. Indian Income Tax Act 1961: Section 80C, 80D, 80CCD, 10(13A) HRA, Standard Deduction, Section 115BAC (Old vs. New Tax Regimes), Section 87A Rebate, and Tax Slabs.
2. Goods and Services Tax (GST): Input Tax Credit (ITC), Section 17(5) Blocked Credits, Composition Scheme, GSTR-1, and GSTR-3B compliance.
3. Financial Advisory & Budgeting: Fixed Deposits, 50/30/20 budgeting rule, portfolio diversification, CTC vs Take-Home breakdowns.
4. Responsible Communication: Stating assumptions, identifying missing user details, and enforcing system safety boundaries.

---

📊 Fine-Tuning & Evaluation Metrics

Parameter	Value
Base Model	Qwen/Qwen2.5-3B-Instruct
Quantization	4-bit NF4 (bitsandbytes)
PEFT Method	QLoRA ($r=8$, $\alpha=16$, dropout=$0.05$, bias=none)
Target Modules	q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
Dataset Size	253 ChatML Q&A examples (202 train / 51 validation, zero data leakage)
Training Hyperparameters	Epochs = 2, Batch Size = 1, Grad Accumulation = 8, Learning Rate = 2e-4, Max Length = 1024
Hardware Used	NVIDIA GeForce RTX 4050 Laptop GPU (6GB VRAM)
Peak VRAM Allocated	2.61 GB / 6.00 GB
Training Duration	596.07 seconds (9.93 minutes)
Final Training Loss	1.2388
Final Validation Loss	0.6851
Benchmark Generation Test	PASS

---

💻 Quickstart Usage Code

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

MODEL_ID = "Qwen/Qwen2.5-3B-Instruct"
ADAPTER_ID = "peterrs23/audit-v-qwen-3b"

# Configure 4-bit NF4 Quantization
bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_use_double_quant=True,
    bnb_4bit_compute_dtype=torch.float16,
)

# Load Tokenizer & Base Model
tokenizer = AutoTokenizer.from_pretrained(ADAPTER_ID, trust_remote_code=True)
base_model = AutoModelForCausalLM.from_pretrained(
    MODEL_ID,
    quantization_config=bnb_config,
    device_map="auto",
    trust_remote_code=True,
)

# Load Fine-Tuned Audit V LoRA Adapter
model = PeftModel.from_pretrained(base_model, ADAPTER_ID)
model.eval()

# Sample Query
messages = [
    {"role": "system", "content": "You are Audit V, an AI financial and tax intelligence assistant."},
    {"role": "user", "content": "What is the maximum deduction allowed under Section 80C?"}
]

prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(prompt, return_tensors="pt").to("cuda")

with torch.no_grad():
    outputs = model.generate(**inputs, max_new_tokens=200, temperature=0.7)

response = tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True)
print(response)

---

⚠️ Important Limitations & Disclaimers

1. Adapter Dependency: This repository contains a PEFT LoRA adapter, not a standalone model. It must be loaded on top of Qwen/Qwen2.5-3B-Instruct.
2. Deterministic Calculation Boundary: The model does not replace Audit V's deterministic Indian Tax Engine for exact tax liability calculations. In production, exact calculations are handled by the tax API, while the LLM provides explanations.
3. Retrieval Integration: In production, this model should be coupled with a RAG vector database (e.g., ChromaDB) for real-time authoritative legal references.
4. Legal & Financial Disclaimer: This model is built for educational and informational assistant workflows. It does not act as a certified Chartered Accountant or legal tax filing authority.