---
license: apache-2.0
base_model: Qwen/Qwen2.5-7B-Instruct
tags:
  - finance
  - financial-qa
  - qlora
  - unsloth
  - qwen2.5
  - quantitative
datasets:
  - TheFinAI/flare-finqa
  - sujet-ai/Sujet-Finance-Instruct-177k
language:
  - en
pipeline_tag: text-generation
library_name: transformers
model-index:
  - name: krishna-finance-7b
    results: []
---

# krishna-finance-7b

A fine-tuned **Qwen2.5-7B-Instruct** model specialized for **financial question answering and quantitative reasoning**. Trained on a combination of financial QA and instruction-following datasets to handle earnings analysis, ratio calculations, financial statement interpretation, and investment reasoning.

## Key Details

| | |
|---|---|
| **Base model** | Qwen/Qwen2.5-7B-Instruct |
| **Method** | QLoRA (4-bit NF4, rank 16, alpha 16) |
| **Library** | Unsloth + TRL SFTTrainer |
| **Datasets** | TheFinAI/flare-finqa (5K) + Sujet-Finance-Instruct-177k (5K) |
| **Total examples** | 10,000 |
| **Hardware** | NVIDIA RTX A5000 (24GB VRAM) on RunPod |
| **Training time** | ~2.75 hours |
| **Parameters trained** | 40.4M of 7.66B (0.53%) |
| **Format** | ChatML (`<\|im_start\|>` / `<\|im_end\|>`) |
| **Output** | Merged 16-bit safetensors |

## Dataset Composition

The training data blends two complementary sources:

- **FinQA** (5,000 examples) — financial question answering requiring numerical reasoning over earnings reports, balance sheets, and financial tables. Teaches the model to extract numbers, perform calculations, and explain financial logic step by step.

- **Sujet Finance Instruct** (5,000 examples) — broad financial instruction data covering investment analysis, market concepts, risk assessment, portfolio management, and financial planning. Gives the model general financial fluency.

## Usage

### Transformers

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("sriksven/krishna-finance-7b")
tokenizer = AutoTokenizer.from_pretrained("sriksven/krishna-finance-7b")

messages = [
    {
        "role": "system",
        "content": "You are a financial analyst. Answer questions about financial data with precise calculations and step-by-step reasoning.",
    },
    {
        "role": "user",
        "content": "A company reported revenue of $120M and cost of goods sold of $75M. Operating expenses were $25M. Calculate the gross margin and operating margin.",
    },
]

inputs = tokenizer.apply_chat_template(messages, return_tensors="pt", add_generation_prompt=True)
outputs = model.generate(inputs, max_new_tokens=512)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
```

### Unsloth (faster inference)

```python
from unsloth import FastLanguageModel

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="sriksven/krishna-finance-7b",
    max_seq_length=2048,
    load_in_4bit=True,
)
FastLanguageModel.for_inference(model)
```

## Example Capabilities

- **Financial ratio calculation** — gross margin, operating margin, ROE, P/E, debt-to-equity
- **Earnings analysis** — interpreting revenue trends, YoY growth, segment performance
- **Financial statement reading** — balance sheet, income statement, cash flow analysis
- **Investment reasoning** — valuation approaches, risk factors, portfolio considerations
- **Quantitative QA** — multi-step numerical reasoning over financial data

## Intended Use

- Financial question answering systems
- Building finance-focused chatbots or copilots
- Quantitative analysis assistants for analysts and students
- Research on domain-specific LLM fine-tuning in finance

## Limitations

- Not a financial advisor — outputs should not be used as investment advice
- Trained on English-language financial data only
- May hallucinate financial figures not present in the input context
- No real-time market data access — knowledge limited to training data patterns
- Not evaluated against established financial NLP benchmarks (FinQA leaderboard, etc.)
- Best results when using the system prompt format matching training

## Training Infrastructure

| | |
|---|---|
| **GPU** | NVIDIA RTX A5000 24GB |
| **Cloud** | RunPod ($0.27/hr) |
| **Framework** | Unsloth 2026.5.2 + TRL + Transformers 5.5.0 |
| **Precision** | BF16 training, 4-bit NF4 base quantization |
| **Optimizer** | AdamW 8-bit |
| **Learning rate** | 2e-4, linear decay |
| **Batch size** | 16 effective (4 per device × 4 accumulation) |
| **Packing** | Enabled |

## Source Code

Training scripts and configs: [github.com/sriksven/LLM-FineTune-Suite](https://github.com/sriksven/LLM-FineTune-Suite)

## License

Apache 2.0