enPurified/dolphin-r1-deepseek-enPurified-openai-messages

enPurified: High-Quality English Prose Collection

This dataset is a pruned version of https://huggingface.co/datasets/mlabonne/dolphin-r1-deepseek

Updated on January 18th to remove as much math and code as possible.

Dataset Summary

The enPurified collection represents a rigorous effort to distill high-value, open-source datasets into their purest form of English prose and reasoning. While the open-source community provides an abundance of datasets optimizing for mathematics, coding, and multilingual capabilities, there is a distinct need for corpora focused exclusively on high-quality English syntactic reasoning and narrative flow.

The goal of this dataset is to serve as a high-fidelity training source for models requiring strong English language fundamentals without the noise of code snippets, mathematical LaTeX blocks, or low-quality web text. This specific subset utilizes a multi-stage heuristic pipeline to prune extraneous data, ensuring only the most coherent and lexically rich examples remain.

All data has been standardized into the OpenAI Messages format for seamless integration into modern training pipelines. For source material involving long-form text (e.g., standardEbooks or LongPage), high-quality LangChain scripts were utilized to chunk narratives into contextually relevant paragraphs.

Dataset Statistics

The application of the enPurified heuristic pipeline resulted in a significant reduction in dataset size, prioritizing quality over quantity.

• Original Row Count: 300,018
• Final Row Count: 38,099
• Reduction Rate: ~87.3% removed
• Content Type: High-quality English reasoning and prose.

The Pruning Pipeline

The curation process involves a strict, ten-stage "Gauntlet" designed to filter out non-prose artifacts. The pipeline operates on the following heuristics:

1. Pre-Normalization

Before analysis, the text is normalized. Solution blocks (<|begin_of_solution|>) are stripped to remove noisy artifacts, and "Thought" tags are standardized to <think>/</think> to preserve chain-of-thought reasoning while making it parseable.

2. The Heuristic Gauntlet

A document must pass all of the following gates to be included in the final dataset:

• Syntax & Ban Gate:

  • Filters out specific coding artifacts (e.g., #include, public static void, console.log).
  • Removes explicit HTML doctypes and Asymptote Vector Graphics ([asy]).
  • Rejects documents containing specific banned substrings indicative of non-prose content.

• Math Gate:

  • Aggressively detects and removes heavy LaTeX formatting.
  • Filters out block equations ($$...$$, \begin{equation}) and dense inline math ($f(x)$).
  • Ensures the text is narrative-driven rather than formula-driven.

• Symbol Density Check:

  • Calculates the ratio of code-like symbols ({, }, =>, ;) relative to the text length.
  • Documents with a symbol density exceeding 4% are rejected to prevent code-mixed text from leaking through.

• Quiz & MCQ Filter:

  • Identifies and removes Multiple Choice Question (MCQ) structures.
  • Detects patterns such as "Option A", "A)", or repeated list markers common in standardized testing datasets.

• Structural Integrity:

  • Short Line Density: Rejects texts where >80% of lines are under 20 characters (indicative of lyrical or list-based data).
  • Repetition: distinct line checks ensure the text is not a repetitive loop or error log.

• Quality Heuristics:

  • Mean Word Length: Validates that the average word length falls between 3.5 and 11 characters, filtering out gibberish or simplified distinct token lists.
  • Length Check: Ensures the document contains sufficient context (minimum 20 words).

• Toxicity Safety:

  • A lightweight filter checking for high-density profanity or NSFW terminology (threshold set at 0.5% of total tokens).

• English Prose Identification:

  • Stopword Density: Requires a minimum density (20%) of common English stopwords (e.g., "the", "is", "of") to ensure the text is grammatically structured English rather than a keyword list.
  • ASCII Check: Enforces a high percentage of ASCII characters to filter out non-English languages.

• MTLD Richness (Lexical Diversity):

  • Calculates the Measure of Textual Lexical Diversity (MTLD).
  • Only documents with an MTLD score ≥ 70.0 are retained, ensuring a high vocabulary range and avoiding repetitive, simple phrasing.

Data Structure

The dataset adheres to the standard OpenAI Messages API format:

{
  "messages": [
    {
      "role": "user",
      "content": "..."
    },
    {
      "role": "assistant",
      "content": "..."
    }
  ]
}

Licensing

This dataset is a derivative work. Users should refer to the license of the original source dataset from which this subset was purified.

https://huggingface.co/datasets/mlabonne/dolphin-r1-deepseek