davidfoss/Synthetic-Causal-Reasoning-50k

🏭 Sovereign Synthetic Reasoning Dataset (400k)

"High-Quality Chain-of-Thought Data at Scale."

📊 Overview

This dataset contains 400,000 synthetic reasoning samples spanning 16 enterprise domains (Finance, Pharma, Legal, Cybersecurity, Supply Chain, etc.).

It was generated using the Sovereign Generator, which produced 1.6 million samples and applied a strict quality filter (Top 25%) to retain only the most logically consistent and complex chains.

Average Quality Score: 90.0 / 100

🚀 Key Features

• —Massive Scale: 400k unique samples for pre-training or fine-tuning.
• —Diverse Styles: Includes Academic (22%), Scientific (20%), QA (20%), Regulatory (7.5%), and Executive Summaries.
• —Strict Logic: Every sample follows the Trigger -> Mechanism -> Outcome causal structure.
• —High Quantification: 100% of samples contain simulated hard data points (metrics, dollars, percentages) to train numerical reasoning.

🏛 Domains Covered

The dataset ensures broad generalization by covering 16 distinct sectors:

• —Critical Infrastructure: Energy, Telecom, Aerospace, Agriculture.
• —Enterprise Risk: Finance, Insurance, Real Estate, Retail.
• —Science & Tech: Pharma, Healthcare, Cybersecurity, Manufacturing.
• —Governance: Legal, Government, Education, Supply Chain.

📂 Data Structure

The dataset follows a standard Instruction Tuning format (prompt / completion), making it compatible with most fine-tuning frameworks immediately.

Each entry trains the model to perform structured causal analysis:

{
  "prompt": "How does patent cliff approaching result in revenue decline in affected product line?",
  "completion": "**Causal Analysis: patent cliff approaching → revenue decline in affected product line**\n\n**1. Trigger Identification**\nThe causal sequence initiates with patent cliff approaching...\n\n**2. Mechanistic Pathway**\nThe underlying mechanism is: generic competition preparation triggers pricing pressure...\n\n**3. Quantitative Evidence**\n**Measured Effect:** typical revenue drop: 70-90% within 12 months\n..."
}

🔗 Related Datasets

This is the Synthetic expansion pack. For ground-truth data extracted from real-world PDFs, see the Sovereign Enterprise Reasoning Dataset (Real) here: 👉 https://huggingface.co/datasets/chkmie/Sovereign-Financial-Aviation-Reasoning

🛠 Methodology

1. 1.Generation: 1.6M raw samples generated via domain-specific prompt engineering.
2. 2.Filtering: Automated scoring based on logical coherence, vocabulary complexity, and structural integrity.
3. 3.Selection: Only the top 25% (Score > 86) were retained to ensure high-quality training signals.

Data Augmentation Strategy

To ensure robust model generalization, the dataset employs multi-style augmentation. Core causal scenarios are presented in various formats (Scientific Report, Executive Brief, QA, Regulatory Filing) to teach the model to recognize the underlying causal logic regardless of the surface-level phrasing.

Generated by chkmie - 2026