---
license: apache-2.0
language:
- en
pipeline_tag: text-generation
tags:
- finetune
- unsloth
- python
- code-generation
- financial-modeling
base_model: Qwen/Qwen2.5-3B-Instruct
pretty_name: FinCode-Reasoning-3B (Full Weights)
library_name: transformers
---

# 📈 FinCode-Reasoning-3B (Full Weights)

[![License](https://shields.io)](https://opensource.org/)
[![Unsloth](https://shields.io)](https://github.com/unslothai/unsloth)
[![Alrithm API](https://shields.io)](https://alrithmapi.vercel.app)

> 🔒 **Verified Ingestion Pipeline:** To stream and verify the cryptographic integrity of structural data assets directly into production AI pipelines, use **[Alrithm API (alrithmapi.vercel.app)](https://vercel.app)** featuring automated row-level Merkle Root integrity checks.

**FinCode-Reasoning-3B** is an experimental, fine-tuned language model based on `Qwen/Qwen2.5-3B-Instruct`. Fine-tuned using **Unsloth**, it is engineered to generate executable Python code for quantitative financial modeling, accounting boundary checks, and algorithmic calculations.

> ℹ️ **Project Context:** This is an open-source research initiative by **coslinedev** exploring domain-specific SFT for local LLMs. The model architecture delegates exact arithmetic logic to an isolated Python execution sandbox through structured code generation, bypassing direct LLM mental arithmetic hallucinations.

---

## 📁 Repository Structure

This repository contains the **full merged model weights** (~6.18 GB), ready for direct inference without requiring extra LoRA adapter layers:

* `model-00001-of-00002.safetensors` & `model-00002-of-00002.safetensors`: Full model weight shards.
* `model.safetensors.index.json`: Index mapping for sharded weight loading.
* `chat_template.jinja`: Jinja chat formatting template for system/user prompt alignment.
* `tokenizer.json` & `tokenizer_config.json`: Tokenizer assets inherited from Qwen2.5.
* `config.json` & `generation_config.json`: Model architecture and generation configuration parameters.

---

## 🛠️ Usage Example

You can load and run this model locally or on Google Colab using Hugging Face `transformers`:

```python
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

MODEL_ID = "coslinedev/Qwen2.5-3B-FinCode-Reasoning-Full"

tokenizer = AutoTokenizer.from_pretrained(MODEL_ID, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    MODEL_ID,
    torch_dtype=torch.float16,
    device_map="auto"
)

prompt = "Write a Python function to calculate Double Declining Balance (DDB) depreciation with a salvage floor constraint."

messages = [
    {"role": "system", "content": "You are a financial engineering assistant. Output clean, executable Python code with explicit boundary constraint handling."},
    {"role": "user", "content": prompt}
]

formatted_prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer([formatted_prompt], return_tensors="pt").to(model.device)

outputs = model.generate(**inputs, max_new_tokens=1024, temperature=0.2)
print(tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True))
```

---

## ⚙️ Key Design Principles

*   **Execution Sandbox First:** Direct numerical output is delegated to the Python interpreter executing the generated script, eliminating calculation hallucinations.
*   **Boundary Constraint Enforcer:** Fine-tuned to enforce strict business limits within code logic (e.g., `depreciation = min(book_value * ddb_rate, max(book_value - salvage, 0.0))`).
*   **Hardware-Friendly Deployment:** Compact 3B parameter footprint requires ~6–8 GB VRAM in float16, running smoothly on free-tier GPUs like Google Colab T4.

---

## 📊 Preliminary Benchmark (n=100)

Evaluated on 100 synthetic quantitative financial tasks (DDB depreciation schedules, Black-Scholes pricing, WACC calculations, Tax Shield bounds, DCF models).

Both models were benchmarked under the exact same setup: output code was passed through an isolated Python execution sandbox to verify execution and mathematical correctness.

| Model Setup | Code Pass Rate (Pass@1) | Valid Syntax Rate | Avg Generation Latency* |
| :--- | :--- | :--- | :--- |
| **FinCode-Reasoning-3B (Ours)** | **98.0%** | **99.5%** | **0.85s** |
| Qwen2.5-3B-Instruct (Base) | 82.0% | 85.0% | 0.82s |

*\*Inference latency measured on a free-tier Google Colab T4 GPU (float16 precision). Does not include sandbox execution overhead.*

---

## ⚠️ Known Limitations & Disclaimers

*   **Sample Size Noise:** Evaluated on n=100 tasks, carrying an estimated statistical error margin of ± 4%.
*   **Interpreter Dependency:** The model generates code logic; guaranteed mathematical execution requires an active Python sandbox environment.
*   **Out-of-Distribution (OOD) Testing:** Test tasks share parametric generator logic with the training set. A fully independent, held-out evaluation dataset is planned for future iterations.

---

## 📬 Feedback & Contact

If you spot technical inaccuracies, unexpected code generation logic, or have ideas for improvement, feel free to contribute!

*   **Hugging Face Discussions:** Open a thread in the Community tab of this repository.
*   **Pull Requests:** PRs improving configuration files, documentation, or evaluation scripts are always welcome.

---

## 📜 License

Licensed under the Apache 2.0 License, aligning with base Qwen2.5 model licensing.