enPurified/SYNTH-enPurified-openai-messages

enPurified Prose Collection

This is derivative work of https://huggingface.co/datasets/PleIAs/SYNTH

Curatorial Objective

The enPurified collection is a curated initiative designed to distill high-value, existing datasets into their purest English prose form. The primary objective is to create a corpus strictly dedicated to high-quality linguistic reasoning and narrative flow, explicitly excluding domain-specific notations that often dilute prose models.

Unlike general-purpose datasets, this collection enforces a strict No Math, No Code, No Low-Quality Text policy. While high-quality datasets exist for programming and mathematics, enPurified targets the nuance of English syntax and vocabulary.

Data Processing Strategy

• Format Standardization: All data is converted into the OpenAI messages format for seamless integration into training pipelines.
• Long-Context Handling: For sources involving extensive text strings (e.g., LongPage or StandardEbooks), a specialized LangChain implementation is utilized to segment narratives into coherent paragraphs, paired with relevant instructional headers in the messages format.
• Synthetic Filtering: For synthetic datasets (e.g., PleIAs/SYNTH), a rigorous heuristic pipeline is applied to strip robotic artifacts and ensure lexical diversity.

This enPurification process trimmed down the Synth dataset from 68,028,365 rows down to 9,323,958 rows of high quality English prose. Math, code, and unnecessary synthetic artifacts were stripped out as much as possible.

---

Heuristic Pruning Pipeline (The Gauntlet)

The dataset is generated via a multi-stage filtering script designed to aggressively prune low-quality or out-of-domain entries. The pipeline operates as follows:

1. Artifact Cleaning & Normalization

Before analysis, raw text undergoes cleaning to remove "robotic garbage" and meta-commentary often found in synthetic chain-of-thought data.

• Removal of Meta-Tags: Strips patterns like [Stream:], Analysis:, NB:, and excessive markdown headers.
• Whitespace Normalization: Collapses distinctive whitespace to standard prose spacing.

2. Structural & Syntax Gating

Entries are rejected if they exhibit structural characteristics typical of code, listicles, or low-effort responses.

• Lazy Thought Detection: Rejects entries where the "reasoning" component is disproportionately short compared to the answer (Logic: < 10% ratio for long answers).
• Bullet Point Density: Discards text where bullet points or numbered lists constitute >25% of the content (or >65% of the thought block) to force narrative prose.
• Line Structure: Filters out "vertical" text where short lines (<30 chars) make up >25% of the document.

3. Domain Exclusion (Math & Code)

A strict set of checks ensures the complete removal of non-prose content.

• Symbol Density: Rejects text where code-specific symbols ({, }, <, >) exceed 3.3% of characters.
• Math Gate: Detects and rejects LaTeX patterns ($$...$$, \begin{...}) and variable assignments.
• Code Syntax: Identifies camelCase variables, function definitions (def, void), and lines ending in code-typical syntax (;, {).
• Banned Substrings: Immediate rejection of strings containing HTML doctypes, specific imports (import matplotlib), or memory addresses.

4. Quality & Safety Heuristics

Final quality control focuses on the linguistic richness of the remaining text.

• English Prose ID: Verifies English stopword density (>14%) and ASCII character ratios (>98%) to filter foreign languages and garbled text.
• Lexical Richness (MTLD): Calculates the Measure of Textual Lexical Diversity. Only texts with an MTLD score of 80+ are retained, ensuring a sophisticated vocabulary.
• MCQ Filter: Removes "Multiple Choice Question" formats.
• Toxicity: Basic keyword filtering for NSFW or toxic content.

---

Output Format

The final dataset is saved in JSONL format using the standard OpenAI schema. Synthetic reasoning (Chain-of-Thought) is preserved but encapsulated within <think> tags inside the assistant's response.

{
  "messages": [
    {
      "role": "user", 
      "content": "User query here..."
    },
    {
      "role": "assistant", 
      "content": "<think>\nSanitized reasoning process...\n</think>\n\nHigh-quality prose answer..."
    }
  ]
}

License

This dataset is a derivative work. Please refer to the license of the original source dataset used for this subset. Users are responsible for complying with the original licensing terms.