Foss Generator

Synthetic Causal Training Data Generator for LLM Fine-Tuning

Part of the Sovereign Causal Graph research project.

Overview

Foss Generator creates high-quality synthetic Chain-of-Thought training data from domain ontologies. Pure Python, no external dependencies, no API calls required.

Performance: ~100,000 samples in ~10 seconds on Apple Silicon M-series.

flowchart LR
    A[Domain Ontology] --> B[Causal Mechanisms]
    B --> C[Template Engine]
    C --> D[Style Variants]
    D --> E[Quality Scoring]
    E --> F[Training Data]

    style A fill:#3b82f6,color:#fff
    style F fill:#10b981,color:#fff

Loading

Features

• 16 Industry Domains: Pharma, Finance, Cybersecurity, Healthcare, Energy, and more
• 12 Output Styles: Academic, executive, regulatory, technical, etc.
• Quality Scoring: Automatic quantification pattern detection
• Multiple Exports: JSONL, HuggingFace-compatible, CSV
• Custom Ontologies: Load your own domain definitions

Installation

git clone https://github.com/DT-Foss/foss-generator.git
cd foss-generator

No dependencies required - pure Python 3.8+.

Usage

Generate samples for a specific domain:

python foss_generator.py --domain pharma --count 10000

Generate samples for all domains:

python foss_generator.py --all-domains --count 5000

Use a custom ontology:

python foss_generator.py --custom my_ontology.json --count 10000

Output formats:

python foss_generator.py --domain finance --count 5000 --format huggingface

Architecture

flowchart TB
    subgraph Input
        O[Domain Ontologies<br/>16 Industries]
        C[Custom Ontology<br/>JSON]
    end

    subgraph Generator
        M[Mechanism Selector]
        T[Template Engine<br/>12 Styles]
        Q[Quality Scorer]
    end

    subgraph Output
        J[JSONL]
        H[HuggingFace]
        V[CSV]
    end

    O --> M
    C --> M
    M --> T
    T --> Q
    Q --> J
    Q --> H
    Q --> V

    style O fill:#3b82f6,color:#fff
    style C fill:#3b82f6,color:#fff
    style Q fill:#f59e0b,color:#fff
    style J fill:#10b981,color:#fff
    style H fill:#10b981,color:#fff
    style V fill:#10b981,color:#fff

Loading

Output Sample

{
  "prompt": "Trace the causal pathway from patent cliff approaching to revenue decline.",
  "completion": "**Causal Analysis: patent cliff approaching → revenue decline**\n\n**1. Trigger Identification**\nThe causal sequence initiates with patent cliff approaching.\n\n**2. Mechanistic Pathway**\nGeneric competition preparation triggers pricing pressure and market share erosion.\n\n**3. Quantitative Evidence**\nTypical revenue drop: 70-90% within 12 months.\n\n**4. Confidence Assessment**\nThis causal relationship demonstrates very_high confidence."
}

Available Domains

Domain	Mechanisms	Description
pharma	18	Drug development, clinical trials, regulatory
finance	18	Risk management, trading, compliance
cybersecurity	15	Threat detection, incident response
healthcare	16	Clinical workflows, patient safety
energy	15	Grid operations, renewable integration
manufacturing	14	Supply chain, quality control
legal	12	Contract analysis, litigation
insurance	14	Underwriting, claims processing
aviation	12	Safety systems, maintenance
telecom	12	Network operations, capacity
biotech	15	Research, development pipelines
mining	12	Extraction, environmental compliance
agriculture	12	Crop management, supply chain
real_estate	12	Valuation, market dynamics
defense	10	Procurement, systems integration
maritime	10	Shipping, port operations

Causal Chain Structure

flowchart LR
    T[Trigger Event] --> M[Mechanism]
    M --> O[Outcome]
    O --> Q[Quantification]

    style T fill:#ef4444,color:#fff
    style M fill:#f59e0b,color:#fff
    style O fill:#10b981,color:#fff
    style Q fill:#8b5cf6,color:#fff

Loading

Each generated sample follows this causal structure:

• Trigger: Initiating condition (e.g., "Phase III trial completion")
• Mechanism: Causal pathway (e.g., "FDA review process activation")
• Outcome: Terminal effect (e.g., "drug approval or rejection")
• Quantification: Measured evidence (e.g., "58% approval rate")

Research

This generator is part of the Sovereign Causal Graph research on LLM-free scientific knowledge extraction. The system achieves:

• Foss Hallucination Gate: 14-step validation pipeline
• Grade A/B extraction rate: 4.6% from raw PDF content
• Validated by: Pieter Wuille (SIPA), Bitcoin Core maintainer

License

MIT License - see LICENSE

Author

David Tom Foss