110 lines
4.2 KiB
Markdown
110 lines
4.2 KiB
Markdown
---
|
|
license: apache-2.0
|
|
tags:
|
|
- text-generation
|
|
- finance
|
|
- economics
|
|
- cpa
|
|
- audcor
|
|
datasets:
|
|
- Josephgflowers/Finance-Instruct-500k
|
|
language:
|
|
- en
|
|
base_model:
|
|
- unsloth/Qwen3-8B-unsloth-bnb-4bit
|
|
pipeline_tag: text-generation
|
|
library_name: transformers
|
|
---
|
|
|
|
# Model Card for CPA-Qwen3-8B-v0
|
|
|
|
**CPA-Qwen3-8B-v0** is a specialized large language model developed by **[AudCor](https://huggingface.co/AudCor)**, designed specifically to assist with Certified Public Accountant (CPA) tasks, regulatory compliance, and financial reasoning.
|
|
|
|
## About AudCor
|
|
**AudCor** is dedicated to building advanced AI tools tailored for CPA firms. Our mission is to enhance the productivity and accuracy of accounting professionals by integrating state-of-the-art language models with deep domain expertise in auditing, tax, and financial reporting.
|
|
|
|
## Key Features
|
|
|
|
* **CPA Expert Persona:** Fine-tuned to adopt the role of a seasoned CPA, prioritizing accuracy, professional skepticism, and strict adherence to GAAP, IFRS, and tax codes.
|
|
* **Domain Specificity:** Optimized for complex financial queries, including audit risk assessment, tax planning strategies, and regulatory interpretation.
|
|
* **Exam-Grade Reasoning:** Capable of handling the rigorous logic required for CPA exam-level problems and real-world accounting scenarios.
|
|
|
|
## CPA Exam Insights
|
|
|
|
Understanding the standards against which this model is benchmarked requires looking at the rigorous CPA Exam scoring process.
|
|
|
|
|
|
|
|
|
|
### The Magic Number: 75
|
|
The passing score or "Gold Standard" for the CPA Exam is **75** on a scale of 0-99.
|
|
* **Crucial Distinction:** This is a *scaled score*, not a percentage correct.
|
|
* **No Curve:** Candidates are measured against a standard of competence, not against each other.
|
|
|
|
### Scoring Mechanics: Multi-Stage Testing (MST)
|
|
The CPA Exam uses an adaptive testing model for Multiple-Choice Questions (MCQs):
|
|
|
|
1. **Testlet 1:** Always "Medium" difficulty.
|
|
2. **Performance Trigger:**
|
|
* *Strong Performance* $\rightarrow$ **Testlet 2: Difficult** (Higher scoring potential).
|
|
* *Weak Performance* $\rightarrow$ **Testlet 2: Medium** (Lower scoring potential).
|
|
|
|
> [!NOTE]
|
|
> Getting a "Difficult" second testlet is desirable because difficult questions are weighted more heavily, raising your potential score ceiling.
|
|
|
|
### Section Scoring Breakdown (2025 Update)
|
|
With the CPA Evolution changes, the scoring weights have shifted:
|
|
|
|
| Section Type | MCQ Weight | TBS (Task-Based Simulations) Weight |
|
|
| :--- | :--- | :--- |
|
|
| **Core Sections (AUD, FAR, REG)** | 50% | 50% |
|
|
| **Discipline Sections** | 50% | 50% |
|
|
|
|
*Data Source: [JobSeekerTools - CPA Passing Score](https://jobseekertools.com/blog/cpa-passing-score)*
|
|
|
|
## Usage
|
|
|
|
This model is compatible with `transformers` and `unsloth`.
|
|
|
|
```python
|
|
from transformers import AutoModelForCausalLM, AutoTokenizer
|
|
|
|
model_name = "AudCor/cpa-qwen3-8b-v0"
|
|
|
|
# Load the model
|
|
model = AutoModelForCausalLM.from_pretrained(
|
|
model_name,
|
|
device_map="auto"
|
|
)
|
|
tokenizer = AutoTokenizer.from_pretrained(model_name)
|
|
|
|
# Specialized System Prompt
|
|
system_prompt = "You are an expert Certified Public Accountant (CPA). Your goal is to provide accurate, professional, and compliant financial advice..."
|
|
|
|
messages = [
|
|
{"role": "system", "content": system_prompt},
|
|
{"role": "user", "content": "Explain the revenue recognition principle under ASC 606."},
|
|
]
|
|
|
|
inputs = tokenizer.apply_chat_template(messages, return_tensors="pt", add_generation_prompt=True).to("cuda")
|
|
outputs = model.generate(inputs, max_new_tokens=256)
|
|
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
|
|
```
|
|
|
|
## Disclaimer
|
|
|
|
**Not Financial Advice.** This model is an artificial intelligence research tool. While trained on high-quality financial data, it is not a licensed CPA and should not be used as a substitute for professional accounting, tax, or legal advice.
|
|
|
|
## Citation
|
|
|
|
```bibtex
|
|
@misc{AudCor/cpa-qwen3-8b-v0,
|
|
author = {AudCor},
|
|
title = {AudCor/cpa-qwen3-8b-v0: A specialized CPA large language model fine-tuned on Josephgflowers/Finance-Instruct-500k},
|
|
year = {2025},
|
|
publisher = {Hugging Face},
|
|
journal = {Hugging Face Model Hub},
|
|
howpublished = {\url{https://huggingface.co/AudCor/cpa-qwen3-8b-v0}}
|
|
}
|
|
```
|