Modelpublic

OsamaAli313/CFO-Agent-14B

sourceHugging Faceapache-2.0updated 6mo agoView on Hugging Face

0likes10downloads

Model Card

CFO-Agent-14B

A fine-tuned language model trained to function as an AI Chief Financial Officer. It provides expert-level financial analysis, forecasting, risk assessment, scenario planning, and executive-level financial communication.

Proof-of-Concept: This version uses Qwen2.5-0.5B-Instruct as the base model for budget-efficient validation. A production version on Qwen2.5-14B-Instruct is planned.

Capabilities

Capability	Description
Financial Statement Analysis	Analyze balance sheets, income statements, and cash flow statements
Revenue & Expense Forecasting	Project future financials with quantified assumptions
Risk Assessment	Identify, quantify, and prioritize financial risks
Scenario Planning	Model best/worst/base case scenarios with financial impact
Budget Optimization	Allocate resources and optimize departmental budgets
Executive Communication	Generate board-level reports, investor updates, and CFO memos
M&A Analysis	Evaluate acquisitions with DCF, multiples, and synergy analysis
Cash Flow Management	Monitor burn rate, runway, and working capital

Quick Start

With PEFT (Recommended)

python

from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel

base_model = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen2.5-0.5B-Instruct",
    device_map="auto",
)
model = PeftModel.from_pretrained(base_model, "OsamaAli313/CFO-Agent-14B")
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-0.5B-Instruct")

messages = [
    {"role": "system", "content": "You are CFO-Agent, an expert Chief Financial Officer AI assistant."},
    {"role": "user", "content": "Analyze this: Revenue $500K, COGS $200K, OpEx $150K. What are our margins?"}
]

text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=512, temperature=0.7, do_sample=True)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

With 4-bit Quantization (Low VRAM)

python

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

bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_use_double_quant=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.bfloat16,
)

base_model = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen2.5-0.5B-Instruct",
    quantization_config=bnb_config,
    device_map="auto",
)
model = PeftModel.from_pretrained(base_model, "OsamaAli313/CFO-Agent-14B")
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-0.5B-Instruct")

Training Details

Dataset

Trained on OsamaAli313/CFO-Agent-14B-Dataset — a curated dataset of 47,481 training samples across three layers:

Layer	Source	Samples	Content
Layer 1	Public HF datasets	~10K	Financial instruction-following (FiQA, Alpaca Finance, FinGPT)
Layer 2	Structured financial data	~5K	Balance sheet analysis, income statement review, earnings calls, risk reports
Layer 3	Synthetic CFO scenarios	~50K	Scenario planning, risk assessment, M&A analysis, budget allocation, executive communication

All data formatted in ChatML with a CFO-Agent system prompt.

Training Configuration

Parameter	Value
Base model	Qwen/Qwen2.5-0.5B-Instruct
Method	QLoRA (SFT)
LoRA rank	64
LoRA alpha	128
Target modules	q, k, v, o, gate, up, down projections
Epochs	3
Batch size	1 (x16 gradient accumulation)
Learning rate	2e-4 (cosine decay)
Max sequence length	512
Optimizer	pagedadamw8bit
Precision	bf16
Hardware	NVIDIA T4 (HF Jobs)
Training time	~3.5 hours

Training Results

Metric	Value
Final training loss	0.1036
Average training loss	0.1706
Mean token accuracy	95.75%
Total steps	938

Training metrics are available on the Trackio dashboard.

Example Prompts

Analyze our income statement: Revenue $500K, COGS $200K, Operating Expenses $150K.
What's our gross margin and operating margin?

We have $100K cash with $20K monthly burn rate. When do we run out of cash?
What cost reduction targets should we set?

Compare two scenarios: Base case 20% revenue growth / 15% margin vs
Worst case 5% growth / 10% margin. What's the year-end profitability impact?

Our revenue grows 10% quarterly starting from $1M.
Project revenue for the next 4 quarters and calculate CAGR.

Limitations

—Proof-of-concept model — built on 0.5B parameter base, limited reasoning depth compared to larger models
—Not financial advice — outputs should be validated by qualified financial professionals
—Training subset — trained on 5,000 of 47,481 available samples due to compute budget
—Context window — trained with max 512 tokens, longer inputs may degrade quality
—No real-time data — cannot access live market data or financial feeds

Roadmap

—[ ] Train on full 47K dataset
—[ ] Scale to Qwen2.5-14B-Instruct base model
—[ ] Add tool-use capabilities (calculator, spreadsheet, API calls)
—[ ] GGUF export for local deployment (Ollama, LM Studio)
—[ ] Evaluation benchmark on financial reasoning tasks

Citation

bibtex

@misc{cfo-agent-14b,
  title={CFO-Agent-14B: An AI Chief Financial Officer},
  author={Osama Ali},
  year={2026},
  url={https://huggingface.co/OsamaAli313/CFO-Agent-14B}
}