README.md

---
license: cc-by-4.0
datasets:
- gtfintechlab/ipo-images
language:
- en
base_model:
- Ultralytics/YOLOv8
pipeline_tag: image-classification
tags:
- finance
- ipo
- image-classification
- ultralytics
- yolo
---

# YOLOv8s — SEC IPO Filing Image Classifier

A fine-tuned [YOLOv8s](https://github.com/ultralytics/ultralytics) model trained to classify images extracted from U.S. IPO registration statements (S-1 and F-1 filings) on [SEC EDGAR](https://www.sec.gov/edgar). This model serves as the initial detection stage in the pipeline used to construct the [gtfintechlab/ipo-images](https://huggingface.co/datasets/gtfintechlab/ipo-images) dataset.

---

## Classes

The model classifies images into 5 categories:

| Label | Description |
|---|---|
| `chart` | Bar charts, line charts, pie charts, org charts, flow charts, etc. |
| `logo` | Company logos and branding marks |
| `map` | Geographic maps |
| `infographic` | Composite visuals combining data, icons, and text |
| `other` | Decorative images, photographs, signatures, and other visuals |

---

## Usage

### Install dependencies
```bash
pip install ultralytics
```

### Run inference
```python
from ultralytics import YOLO

model = YOLO("<path/to/model.pt>")

# Single image
results = model("path/to/image.png")
print(results[0].probs.top1)        # top class index
print(results[0].names)             # class name mapping

# With a confidence threshold
results = model("path/to/image.png", conf=0.5)

# Batch inference
results = model(["image1.png", "image2.png", "image3.png"])
for r in results:
    print(r.probs.top1cls, r.names[r.probs.top1])
```

### Get the predicted label as a string
```python
result = model("image.png")[0]
label = result.names[result.probs.top1]
print(label)  # e.g. "chart"
```

---

## Relation to the IPO Image Dataset

This model is the **first stage** of the classification pipeline used to build the [`gtfintechlab/ipo-images`](https://huggingface.co/datasets/gtfintechlab/ipo-images) dataset — a large-scale collection of 76,000+ labeled images from SEC IPO filings spanning 1994–2026.

The pipeline works as follows:

1. **This model** generates an initial prediction (`initial_yolo_prediction`) for each image
2. An **ensemble of 8 Vision-Language Models** verifies the prediction, producing a consensus score (`llm_yolo_verification_score`) and per-model votes (`llm_yolo_verification_votes`)
3. The final `label` in the dataset reflects this verified output

---

## Citation

If you use this model in your work, please cite:
```bibtex
@misc{galarnyk2026ipomine,
  title  = {IPO-Mine: A Toolkit and Dataset for Section-Structured Analysis of Long, Multimodal IPO Documents},
  author = {Galarnyk, Michael and Lohani, Siddharth and Nandi, Sagnik and Patel, Aman and Kannan, Vidhyakshaya and Banerjee, Prasun and Routu, Rutwik and Ye, Liqin and Hiray, Arnav and Somani, Siddhartha and Chava, Sudheer},
  year   = {2026},
  url    = {https://huggingface.co/datasets/gtfintechlab/ipo-images},
  note   = {Preprint/Working Paper}
}
```