enPurified/l4-enPurified-openai-messages

enPurified: L4 (High-Quality English Prose)

The original dataset https://huggingface.co/datasets/glaiveai/l4 was reduced from ~1,280,000 rows dowwn to 53,574 rows of high quality English prose.

Dataset Description

The enPurified collection represents a rigorous effort to distill existing high-value datasets into their purest English prose forms. The primary objective of this collection is to facilitate the training of language models on high-quality, coherent, and syntactically sound English text, explicitly excluding code generation, mathematical solving, and low-resource foreign languages.

While specialized datasets exist for coding and mathematics, the enPurified suite focuses on narrative flow, reasoning, and linguistic richness. This version derives from glaiveai/l4, processed through a strict heuristic pipeline to retain only high-quality English interactions.

Most datasets in this collection are converted into the standard OpenAI messages format to streamline Supervised Fine-Tuning (SFT). For source data involving long-context strings (such as the LongPage or StandardEbooks subsets), a high-precision LangChain script is utilized to chunk narratives into coherent paragraphs with relevant instruction pairings.

Processing Methodology: The Pruning Pipeline

The purification process involves a "Gauntlet" of heuristic filters designed to aggressively cull low-quality data, "AI slop," and non-prose content. The pipeline operates on a rejection-sampling basis: if a document fails a single gate, it is discarded.

The filtering stages for this dataset are as follows:

1. Pre-Processing & Normalization

Before analysis, specific tokens (such as thought markers or solution blocks) are normalized. The system parses the conversational history to ensure the total conversation length and individual message lengths meet minimum viability thresholds (e.g., User messages < 7 words or Assistant messages < 45 words are often discarded to prevent low-context noise).

2. Code and Math Exclusion

To ensure the dataset remains focused on prose:

• Symbol Density Check: Calculates the ratio of code-specific symbols ({, }, ;, //, etc.). Documents exceeding a 4% symbol density or containing high concentrations of code-like lines are rejected.
• Math Gate: Detects and removes LaTeX formatting (e.g., \begin{equation}), ASCII matrices, and heavy backslash usage typical of mathematical notation.
• Syntax Filtering: Explicitly bans programming artifacts including HTML tags, code blocks, and language-specific keywords (e.g., public static void, import os, std::, console.log).

3. Structural Integrity & Data Hygiene

• JSON/Data Dump Filter: Rejects documents that appear to be raw JSON objects or key-value pair dumps rather than natural language.
• Quiz/MCQ Filter: Identifies and removes Multiple Choice Questions, preventing the model from learning distinct "A/B/C/D" exam-style formatting which degrades prose generation capabilities.
• Repetition Detection: Analyzes unique line counts and n-grams to detect loops or highly repetitive text.
• Broken Syntax: Filters out content with extremely short lines or malformed formatting indicative of scraping errors.

4. "AI Slop" and Stylistic Filtering

A significant portion of the pipeline is dedicated to removing the stylistic artifacts of RLHF-aligned models:

• Refusal & Moralizing: Removes standard refusal phrases ("I cannot," "As an AI") and preachy moralizing blocks ("It is important to note," "fostering a culture of").
• Conversational Fillers: Strips out vacuous conversational padding such as "Hope this helps," "Here is a list," or "Let me know if you need anything else."
• Meta-Commentary: Removes self-referential text where the model discusses its own analysis ("Based on the text provided...").

5. Safety & Toxicity

The pipeline employs a zero-tolerance policy (0.001 threshold) for:

• Toxicity: Aggressive filtering of slurs, hate speech, and profanity.
• NSFW/Erotica: A dual-layer check detects specific fetish terminology and narrative erotica patterns.
• Jailbreak Triggers: Scans for prompt injection attempts (e.g., "DAN," "Ignore previous instructions").

6. Linguistic Richness (MTLD)

To ensure the remaining English text is of high quality, the pipeline calculates the Measure of Textual Lexical Diversity (MTLD). Only documents that demonstrate sufficient lexical variety (MTLD score >= 62.0) and a healthy ratio of standard English stopwords are retained.

Dataset Structure

The dataset is formatted as a JSONL file, adhering to the OpenAI messages schema.

{
  "messages": [
    {
      "role": "user",
      "content": "Explain the concept of entropy in thermodynamics using an analogy."
    },
    {
      "role": "assistant",
      "content": "Imagine a pristine sandcastle sitting on a beach. This sandcastle represents low entropy..."
    }
  ]
}

License

This dataset is a derivative work. Users should refer to the license of the original dataset (glaiveai/l4) for terms of use, attribution requirements, and restrictions.