---
language:
- en
license: odc-by
task_categories:
- text-generation
- translation
tags:
- text-simplification
- information-extraction
- readability
- fine-web
configs:
- config_name: sample_k10
  data_files:
  - path:
    - sample_k10.jsonl.zst
    split: train
- config_name: sample_k100
  data_files:
  - path:
    - sample_k100.jsonl.zst
    split: train
- config_name: sample_k1000
  data_files:
  - path:
    - sample_k1000.jsonl.zst
    split: train
- config_name: sample_k10000
  data_files:
  - path:
    - sample_k10000.jsonl.zst
    split: train
- config_name: sample_k100000
  default: true
  data_files:
  - path:
    - sample_k100000.jsonl.zst
    split: train
- config_name: sample_k20
  data_files:
  - path:
    - sample_k20.jsonl.zst
    split: train
- config_name: sample_k200
  data_files:
  - path:
    - sample_k200.jsonl.zst
    split: train
- config_name: sample_k2000
  data_files:
  - path:
    - sample_k2000.jsonl.zst
    split: train
- config_name: sample_k20000
  data_files:
  - path:
    - sample_k20000.jsonl.zst
    split: train
- config_name: sample_k50
  data_files:
  - path:
    - sample_k50.jsonl.zst
    split: train
- config_name: sample_k500
  data_files:
  - path:
    - sample_k500.jsonl.zst
    split: train
- config_name: sample_k5000
  data_files:
  - path:
    - sample_k5000.jsonl.zst
    split: train
- config_name: sample_k50000
  data_files:
  - path:
    - sample_k50000.jsonl.zst
    split: train
---
# Low Readability Text Dataset

This dataset consists of high-complexity English web text with an estimated readability at or above the **U.S. Grade 12 level**. The content typically features advanced, highly technical prose or verbose syntactical structures, making it well-suited for researching complex language understanding and automation.

### Primary Use Cases

* **Text Simplification:** Training and evaluating models to translate complex text into plain English.
* **Information Extraction (IE):** Benchmarking NLP systems on dense, complex text structures that general human readers find difficult to parse.
* **Readability Assessment:** Training classifiers to recognize specialized, collegiate, or advanced reading levels.

## Dataset Creation & Methodology

### 1. Extraction & Chunking

* Text was ingested from major web corpuses and broken into chunks of **1024 GPT-4 tokens**, featuring a **10-token overlap** between consecutive segments to maintain contextual continuity.
* **50,000 chunks** per source dataset were initially isolated where the GIST-small-readability classifier scored the text at ≥ 12 (equivalent to U.S. high school senior/university freshman level or higher).

### 2. Quality Filtering

* The initial pool was filtered using the [agentlans/GIST-all-MiniLM-L6-v2-quality-v3](https://huggingface.co/agentlans/GIST-all-MiniLM-L6-v2-quality-v3) model. Only chunks with a quality score **> 1.0** were retained.
* In addition, Markov chain-generated text were filtered out using [agentlans/GIST-small-markov-slop-detector](https://huggingface.co/agentlans/GIST-small-markov-slop-detector).

### 3. Clustering

* Post-filtering, data reduction and organization were performed via **Agglomerative Clustering** utilizing [MongoDB/mdbr-leaf-mt](https://huggingface.co/MongoDB/mdbr-leaf-mt) embeddings to group similar semantic contexts.

## Data Splits & Source Composition

The current primary split is sample_k100000. The distribution of source rows is outlined below:

| Source | Rows |
|-----|------:|
| [openbmb/Ultra-FineWeb en](https://huggingface.co/datasets/openbmb/Ultra-FineWeb) | 28842 |
| [HuggingFaceFW/fineweb-edu](https://huggingface.co/datasets/HuggingFaceFW/fineweb-edu) | 25499 |
| [epfml/FineWeb-HQ](https://huggingface.co/datasets/epfml/FineWeb-HQ) | 20403 |
| [Zyphra/Zyda-2](https://huggingface.co/datasets/Zyphra/Zyda-2) | 19156 |
| [EssentialAI/essential-web-v1.0](https://huggingface.co/datasets/EssentialAI/essential-web-v1.0) | 6092 |
| [allenai/dolma3_pool](https://huggingface.co/datasets/allenai/dolma3_pool) * | 8 |

* Low retention due to high prevalence of spam

## Dataset Structure

### Features

* text *(String)*: The 1024-token text block.
* grade *(Float)*: The assigned reading grade level (minimum 12.0).
* source *(String)*: The originating dataset identifier.

### Example Row

```json
{
  "text": "the patents , Galium-Arsenide 935 nm 730 nm?270-275nm? 310-320nm and 760nm (H2O) 230-240nm (N2) 155-160nm (o2) Water vapor-nitrogen absorption at CO(2) laser frequencies. Peterson JC, Thomas ME, Nordstrom RJ, Damon EK, Long RK. \"...a series of pressure-broadened water vapor absorption measurements at 27 CO(2) laser frequencies between 935 cm(-1) and 1082 cm(-1)\" D20 (Deterium heavy water) - shift H2O from 760nm to 1000nm Balmer series or Balmer lines in atomic physics, is the designation of one of a set of six different named series describing the spectral line emissions of the hydrogen atom...",
  "grade": 12.0,
  "source": "Zyphra/Zyda-2"
}

```

## Limitations & Considerations

⚠️ **Usage Warnings:**
* **Boundary Fragmentation:** Because the dataset relies on rigid token-based windowing (1024 tokens), text blocks may start or end mid-sentence or mid-paragraph.
* **Domain Generality:** The texts represent a broad, general-purpose web crawl. While complex, the text does not strictly represent deep, peer-reviewed domains like academic law or professional medicine.
* **Web Noise & Accuracy:** The corpus stems from unverified internet pages. It contains messy syntax, optical character recognition/scraping artifacts, and may include scientifically or historically inaccurate claims.

## Licence

This dataset is made available under the **Open Data Commons Attribution License (ODC-By v1.0)**.