Datasets:

agentlans
/

garbled-text

license: cc-by-4.0
configs:
  - config_name: madlibs
    data_files:
      - path:
          - madlibs.jsonl.zst
        split: train
  - config_name: replacement
    data_files:
      - path:
          - replacement.jsonl.zst
        split: train
  - config_name: scrambled
    data_files:
      - path:
          - scrambled.jsonl.zst
        split: train
task_categories:
  - text-classification
language:
  - en

Garbled Text Dataset

This dataset contains superficially meaningful English text that entirely lacks global coherence and meaning. While individual sentences in the scrambled and replacement subsets are grammatically valid, when combined, they do not form a cohesive narrative or logical text. This dataset is designed to train models on adversarial text classification, natural language inference (NLI), and coherence detection.

Dataset Summary

The dataset is derived from the sample_k10000 split of the agentlans/high-quality-text-long dataset. Each row in this dataset directly maps to the corresponding row of the original source dataset, processed through three randomization algorithms.

Supported Tasks and Leaderboards

Adversarial Text Classification / Coherence Detection: The dataset can be used to train classifiers to distinguish between naturally cohesive text and algorithmic/artificial gibberish.

Dataset Structure

Data Subsets

The dataset is divided into three subsets based on the transformation algorithm applied:

Split Name	Description	Linguistic Characteristics
`scrambled`	Sentences from the original text are shuffled into a random order.	Locally grammatical; lacks global chronological or logical coherence.
`madlibs`	Nouns and verbs within the text are randomly permuted across the document.	Destroys syntax and local semantics; grammatically chaotic.
`replacement`	One-third (1/3) of the sentences in the original text are randomly replaced with sentences from agentlans/high-quality-english-sentences.	Disrupted narrative flow; contains sudden, completely unrelated topics.

Sentence tokenization and Part-of-Speech (PoS) tagging for the transformations were performed using spaCy's en_core_web_sm pipeline.

Limitations

Artificial Patterns: The dataset relies on well-defined, randomized rule-based algorithms. Models trained heavily on this data may overfit to these specific algorithmic artifacts.
Scope of LLM Failures: This dataset does not capture all common Large Language Model (LLM) degradation modes, such as repetitive loops, hallucinations, or subtle factual contradictions.
Not for Pre-training:

Do not use this dataset for standard language modelling or pre-training production LLMs, as it will degrade their ability to generate coherent text.

Additional Information

Licensing

This dataset is licensed under the Creative Commons Attribution 4.0 International (CC BY 4.0) license.

Citation

If you publish work or release a model based on this dataset, please cite it using the following format:

@misc{garbled_text_dataset,
  author = {agentlans},
  title = {Garbled Text Dataset},
  year = {2026},
  publisher = {Hugging Face},
  journal = {Hugging Face Datasets},
  howpublished = {\url{https://huggingface.co/datasets/agentlans/garbled-text}}
}