README.md

---
dataset_info:
  features:
  - name: cik
    dtype: string
  - name: company
    dtype: string
  - name: year
    dtype: int64
  - name: filing_date
    dtype: string
  - name: sic
    dtype: string
  - name: state_of_inc
    dtype: string
  - name: filing_html_index
    dtype: string
  - name: accession_number
    dtype: string
  - name: table_image
    dtype: image
  - name: table_body
    dtype: string
  - name: executives
    dtype: string
  - name: sct_probability
    dtype: float64
  splits:
  - name: train
    num_bytes: 842796643
    num_examples: 7473
  download_size: 764628352
  dataset_size: 842796643
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
license: mit
task_categories:
- table-to-text
language:
- en
tags:
- finance
- sec
- executive-compensation
- def14a
- proxy-statements
pretty_name: SEC Executive Compensation
size_categories:
- 1K<n<10K
---

# SEC Executive Compensation Dataset

> [!CAUTION]
> ## 🚧 DATASET UNDER CONSTRUCTION 🚧
> 
> This dataset is actively being developed and expanded. The current version contains **~2,000 documents** out of a target of **100,000+ SEC filings** (2005-2022).
> 
> **What to expect:**
> - Data may contain errors or inconsistencies
> - Schema and fields may change
> - More records will be added regularly
> - Statistics will be updated as processing continues
> 
> **Use at your own risk for research purposes only.**

🔗 **Pipeline**: [github.com/pierpierpy/Execcomp-AI](https://github.com/pierpierpy/Execcomp-AI)

Structured executive compensation data extracted from SEC DEF 14A proxy statements using AI.

<details>
<summary><b>📊 Dataset Statistics</b> (click to expand)</summary>

![Pipeline Stats](docs/stats_pipeline.png)

![Document Breakdown](docs/chart_pipeline.png)

![Tables by Year](docs/chart_by_year.png)

</details>

---

<details>
<summary><b>🎯 SCT Probability & Quality</b> (click to expand)</summary>

### Why `sct_probability`?

The main VLM (Qwen3-32B) sometimes:
1. **False positives**: Classifies non-SCT tables as SCT (e.g., Director Compensation tables)
2. **Duplicates**: Some documents have multiple tables classified as SCT when only one is the real Summary Compensation Table

A **fine-tuned binary classifier** (Qwen3-VL-4B) scores each table with a probability (0-1) to help filter these cases.

![Probability Stats](docs/stats_probability.png)

![Probability Distribution](docs/chart_probability.png)

> 💡 **Recommendation**: Filter by `sct_probability >= 0.7` to get high-confidence records only.

</details>

---

<details>
<summary><b>💰 Compensation Statistics</b> (click to expand)</summary>

![Compensation Stats](docs/stats_compensation.png)

![Compensation Breakdown](docs/stats_breakdown.png)

### 🏆 Top 10 Highest Paid Executives

![Top 10](docs/stats_top10.png)

### Compensation Distribution

![Compensation Distribution](docs/chart_distribution.png)

### Trends Over Time

![Compensation Trends](docs/chart_trends.png)

</details>

---

## 📋 Dataset Description

This dataset contains **Summary Compensation Tables** extracted from SEC filings, with:
- Original table images
- HTML table structure  
- Structured JSON with executive compensation details
- **SCT probability score** from a fine-tuned binary classifier (to filter false positives)

> 💡 **Tip**: Filter by `sct_probability >= 0.7` to get high-confidence SCT records only.

## Fields

| Field | Type | Description |
|-------|------|-------------|
| `cik` | string | SEC Central Index Key |
| `company` | string | Company name |
| `year` | int | Filing year |
| `filing_date` | string | SEC filing date |
| `sic` | string | Standard Industrial Classification code |
| `state_of_inc` | string | State of incorporation |
| `filing_html_index` | string | Link to SEC filing |
| `accession_number` | string | SEC accession number |
| `table_image` | image | Extracted table image |
| `table_body` | string | HTML table content |
| `executives` | string | JSON array of executive compensation |
| `sct_probability` | float | Probability (0-1) that this is a real SCT (from fine-tuned classifier) |

## Executive Schema

```json
{
  "name": "John Smith",
  "title": "CEO",
  "fiscal_year": 2023,
  "salary": 500000,
  "bonus": 100000,
  "stock_awards": 2000000,
  "option_awards": 500000,
  "non_equity_incentive": 300000,
  "change_in_pension": 50000,
  "other_compensation": 25000,
  "total": 3475000
}
```

---

## 🚀 Quick Start

```python
from datasets import load_dataset
import json

# Load dataset
ds = load_dataset("pierjoe/execcomp-ai")

# Filter by SCT probability (recommended to reduce false positives)
ds_filtered = ds.filter(lambda x: x["sct_probability"] >= 0.7)
print(f"Filtered: {len(ds['train'])} → {len(ds_filtered['train'])} records")

# View first record
print(ds["train"][0])

# Parse executives JSON
record = ds["train"][0]
executives = json.loads(record["executives"])
for exec in executives:
    print(f"{exec['name']} ({exec['title']}): ${exec['total']:,}")
```

### Analyze with Pandas

```python
import pandas as pd
import json

# Convert to DataFrame
df = ds["train"].to_pandas()

# Parse all executives
all_execs = []
for _, row in df.iterrows():
    for exec_data in json.loads(row['executives']):
        exec_data['company'] = row['company']
        exec_data['year'] = row['year']
        all_execs.append(exec_data)

exec_df = pd.DataFrame(all_execs)

# Average compensation by year
print(exec_df.groupby('year')['total'].mean())

# Top 10 highest paid
print(exec_df.nlargest(10, 'total')[['name', 'company', 'year', 'total']])
```

### View Table Image

```python
# Display table image
record = ds["train"][0]
record["table_image"]  # PIL Image object
```

---

## 🔧 Source & Methodology

Data extracted from [SEC EDGAR](https://www.sec.gov/edgar) DEF 14A filings using:
- **[MinerU](https://github.com/opendatalab/MinerU)** for PDF table extraction
- **[Qwen3-VL-32B](https://huggingface.co/Qwen/Qwen3-VL-32B-Instruct)** for classification and extraction

### Pipeline Steps
1. Download DEF 14A PDFs from SEC EDGAR
2. Extract tables with MinerU (VLM-based)
3. Classify tables to identify Summary Compensation Tables
4. Merge tables split across pages
5. Extract structured compensation data with VLM

---

## 📜 License

MIT License

---

## 📖 Citation

If you use this dataset in your research, please cite:

```bibtex
@dataset{execcomp_ai_2026,
  author       = {Di Pasquale, Pier Paolo},
  title        = {SEC Executive Compensation Dataset},
  year         = {2026},
  publisher    = {Hugging Face},
  url          = {https://huggingface.co/datasets/pierjoe/execcomp-ai-sample},
  note         = {AI-extracted executive compensation data from SEC DEF 14A filings (2005-2022)}
}
```

Or in text format:

> Di Pasquale, P. P. (2026). *SEC Executive Compensation Dataset*. Hugging Face. https://huggingface.co/datasets/pierjoe/execcomp-ai-sample

## 🔗 Links

- **GitHub**: [github.com/pierpierpy/Execcomp-AI](https://github.com/pierpierpy/Execcomp-AI)
- **SEC EDGAR**: [sec.gov/edgar](https://www.sec.gov/edgar)