fin-llm-qwen3.5-9b
Model Summary
fin-llm-qwen3.5-9b is a finance-specialized fine-tuned language model based on Qwen3.5-9B. It is designed to assist with structured financial analysis tasks such as financial statement interpretation, ratio analysis, margin analysis, liquidity and solvency calculations, valuation reasoning, finance math, and analyst-style explanation generation.
The model was fine-tuned to produce concise, formula-driven financial answers. Its preferred response style is:
- State the formula.
- Substitute the given values.
- Compute the result.
- Provide a clear final answer.
- State when required data is missing instead of inventing assumptions.
Fine-tuning is intended to customize a pretrained model’s behavior, inject domain-specific task patterns, and optimize it for specialized use cases.
Intended Use
This model is intended for educational, analytical, and workflow-assistance use in finance-related tasks, including:
- Income statement, balance sheet, and cash flow statement analysis
- Gross margin, operating margin, net margin, ROA, ROE, ROIC, current ratio, quick ratio, debt ratios, and working capital metrics
- DCF, WACC, terminal value, comparable company analysis, and valuation walkthroughs
- NPV, IRR, annuities, bond pricing, duration, convexity, CAPM, Sharpe ratio, and portfolio math
- Summarizing and explaining financial concepts
- Generating analyst-style responses with formulas and step-by-step calculations
- Checking whether enough information is available to answer a finance question
Out-of-Scope Use
This model should not be used as a standalone source for:
- Investment advice
- Trading signals
- Legal, tax, or accounting advice
- Audited financial conclusions
- High-stakes financial decision-making without human review
- Predictions of stock prices, market movements, creditworthiness, or bankruptcy risk without external validated data
The model may produce incorrect calculations, outdated assumptions, or overconfident explanations. Human verification is required.
Training Method
The model was fine-tuned using supervised fine-tuning using LoRA, which trains a small set of adapter weights while keeping the base model mostly frozen, while QLoRA uses 4-bit quantization to reduce memory usage.
Recommended LoRA/QLoRA training practices include using learning rates around 2e-4 for normal LoRA/QLoRA fine-tuning, training for roughly 1-3 epochs, and monitoring for overfitting.
Training Data
The model was trained on finance-focused instruction examples covering:
- Formula-based finance calculations
- Financial ratio calculations
- Accounting and financial statement questions
- Analyst-style explanations
- “Use only provided data” instruction-following
- Missing-data refusal behavior
Example training style:
1{
2 "messages": [
3 {
4 "role": "system",
5 "content": "You are a finance analyst assistant. Use only the provided data. Do not invent numbers. Show formulas when calculating. If required data is missing, state what is missing."
6 },
7 {
8 "role": "user",
9 "content": "Revenue is $1,200.00 million and COGS is $480.00 million. Calculate gross margin."
10 },
11 {
12 "role": "assistant",
13 "content": "Formula: Gross margin = Revenue - COGS\nSubstitution: 1,200 - 480 = 720\nFinal answer: Gross margin is $720.00 million."
14 }
15 ]
16}
