---
dataset_info:
  features:
  - name: text
    dtype: string
  - name: meta
    struct:
    - name: ticker
      dtype: string
    - name: year
      dtype: string
    - name: source
      dtype: string
  - name: tokens
    dtype: int64
  - name: score
    dtype: int64
  splits:
  - name: train
    num_bytes: 17816672
    num_examples: 4643
  download_size: 6287329
  dataset_size: 17816672
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
license: mit
task_categories:
- text-generation
- fill-mask
language:
- en
tags:
- finance
- sec
- 10-k
- raw-text
pretty_name: 'Arb Agent: Raw 10-K Data'
size_categories:
- 1K<n<10K
---

# Arb-Agent Raw Data

This is the main text corpus used to train the QuantOxide Reasoning Agent.
It contains clean, semantic text chunks extracted from the 10-K filings of the top 50 S&P 500 companies.

## The Parsing Logic

Parsing SEC filings is unbelievably difficult due to inconsistent HTML, broken table tags, and "incorporation by reference." This dataset was created using a unique parsing technique:

Instead of relying on broken regex headers, the parser scores chunks based on "Financial Density" (presence of metrics, reasoning keywords, dates, or lack of boilerplate language). Additionally, all HTML tables are parsed and converted into Markdown to preserve column alignment for LLM tokenization. 


## Dataset Contents

- Years: 2022, 2023, 2024
- Sections:
  - Item 1A: Risk Factors (Predictive signals)
  - Item 7: Management's Discussion & Analysis (Causal signals)
  - Item 7A: Quantitative Market Risk (Exposure signals)

## Usage

This dataset is ideal for:

Pre-training financial Language Models (Continued Pre-training). \
Testing RAG (Retrieval Augmented Generation) chunking strategies. \
Financial NLP research (Sentiment Analysis on Risk Factors).

```python
from datasets import load_dataset

dataset = load_dataset("YourUsername/quantoxide-raw-10k")
```