Mindmodel.io is an AI-driven cognitive assessment application designed to democratize access to personalized cognitive insights. The platform provides low-cost, accessible alternatives to traditional cognitive assessments, helping users understand their unique cognitive profiles without relying on conventional IQ metrics.

Our platform leverages a custom fine-tuned AI model specifically trained on cognitive science research. The model was created through a structured two-phase training approach to ensure deep domain expertise and accurate cognitive assessment capabilities.

• Interactive cognitive tests measuring memory, attention, processing speed, and executive functions
• Personalized cognitive profiles highlighting individual strengths and areas for development
• AI-driven analysis providing tailored strategies and recommendations
• Progress tracking to monitor cognitive development over time

Mindmodel.io empowers users with constructive, non-judgmental feedback and actionable recommendations for academic, professional, and personal optimization. By leveraging insights about different learning styles and cognitive capabilities, users can improve their learning, productivity, and problem-solving approaches.

Our model training follows a structured two-phase approach:

In this initial phase, we adapted the DeepSeek-R1-Distill-Llama-8B model to the cognitive science domain by exposing it to a large corpus of cognitive science texts. This phase helped the model learn domain-specific vocabulary, concepts, and patterns.

Key aspects:

• Sequential processing of research papers to maintain context
• 4-bit quantization and LoRA for memory-efficient training
• Regular checkpointing and detailed metric monitoring
• Resulted in our domain-adapted model: George-API/DeepSeek-Cognitive-Science

Building on Phase 1, we performed supervised fine-tuning to teach the model to understand complex multidiscinplinary question answer pairing on cognitive science topics.

Key aspects:

• Used instruction-response pairs covering various cognitive science topics
• Lower learning rate (1e-5) for stable fine-tuning
• Focus on question answering and explanation capabilities
• Resulted in our final model: George-API/DeepSeek-Cognitive-Science-SFT

Our training data was processed through a comprehensive pipeline:

1. Collection and Extraction: Raw research papers processed from PDFs to text
2. Cleaning and ID Assignment: Structured text with unique identifiers
3. Segmentation: Papers divided into manageable chunks (1,500-2,000 tokens)
4. Metadata and Tag Assignment: Added detailed cognitive science tags and relationships
5. Final Optimization: Converted to training-ready JSONL format

The pipeline created a hierarchical knowledge graph with:

• Primary categories: Cognitive Domains, Thinking Styles, Contexts, Neurodevelopmental Conditions, Applied Cognition
• Complex relationships between concepts (hierarchical, directional, bidirectional)
• Integration of multiple disciplinary perspectives

The training data focuses on interdisciplinary cognitive science research with emphasis on:

• Cognitive processes and mental models
• Neuroscience and brain function
• Psychological frameworks
• Educational applications
• Neurodiversity understanding
• Workplace/organizational applications

1. Deep Domain Expertise

   • Understanding complex cognitive and neural mechanisms
   • Integrating multiple theoretical frameworks
   • Maintaining scientific accuracy and rigor
   • Handling specialized terminology correctly

2. Interdisciplinary Integration

   • Connecting insights across disciplines:
     • Cognitive Science ↔ Neuroscience
     • Psychology ↔ Education
     • Theory ↔ Practice
   • Understanding cross-domain implications
   • Synthesizing research findings

3. Nuanced Personalized Insight

   • Distinguishing between similar concepts
   • Understanding context-dependent interpretations
   • Recognizing subtle differences in:
     • Cognitive processes
     • Learning approaches
     • Individual differences
     • Neurodiversity manifestations

Based on metadata keywords:

1. Cognitive Processes

   • Attention types (sustained, selective, divided)
   • Memory systems
   • Executive functions
   • Problem-solving approaches
   • Decision-making processes

2. Individual Differences

   • Learning styles
   • Cognitive diversity
   • Neurodevelopmental variations
   • Personal strengths and challenges

3. Practical Applications

   • Educational strategies
   • Workplace accommodations
   • Support systems
   • Assessment approaches
   • Intervention methods

The fine-tuned model should demonstrate:

1. Technical Accuracy

   • Correct use of scientific terminology
   • Accurate representation of research findings
   • Proper citation and reference understanding

2. Contextual Understanding

   • Appropriate application of theories to specific situations
   • Recognition of individual differences
   • Awareness of contextual factors

3. Practical Insight

   • Actionable recommendations
   • Evidence-based solutions
   • Real-world application understanding

1. Data Quality

   • Scientific papers and research findings
   • Peer-reviewed content
   • Current theoretical frameworks
   • Evidence-based practices

2. Balance

   • Theory vs. practical application
   • Different disciplinary perspectives
   • Various cognitive domains
   • Diverse population considerations

3. Ethical Considerations

   • Neurodiversity-affirming approach
   • Individual difference respect
   • Cultural sensitivity
   • Evidence-based recommendations

• Training data has been cleaned and structured
• Categories have been removed for more flexible learning
• Focus on maintaining scientific rigor while ensuring practical applicability
• Emphasis on integrating multiple perspectives and approaches

flowchart TD classDef dataProcessing fill:#d4f1f9,stroke:#05a,stroke-width:1px classDef phase1 fill:#d5f5d5,stroke:#070,stroke-width:1px classDef phase2Prep fill:#fff2cc,stroke:#d6b600,stroke-width:1px classDef phase2SFT fill:#ffe0cc,stroke:#d66b00,stroke-width:1px classDef webApp fill:#e1d4f9,stroke:#5503a9,stroke-width:1px

subgraph DP ["Data Processing Pipeline"]
    A[Research Paper Collection] --> B[PDF Extraction & Cleaning]
    B --> C[Text Segmentation & Chunking]
    C --> D[Metadata & Tag Assignment]
    D --> E[Knowledge Graph Creation]
    E --> F[Training-Ready JSONL Format]
end

subgraph P1 ["Phase 1: Domain Adaptation"]
    G[DeepSeek-R1-Distill-Llama-8B Base Model] --> H[Unsupervised Training with Cognitive Science Corpus]
    H --> I[Domain-Adapted Model:<br>George-API/DeepSeek-Cognitive-Science]
end

subgraph P2P ["Phase 2 Preparation"]
    J[Question Collection for Cognitive Science Topics] --> K[Summarization Model Generates Expert Responses]
    K --> L[Format as Instruction-Response Pairs]
end

subgraph P2S ["Phase 2: Supervised Fine-Tuning (SFT) with LoRA"]
    SP[" "]:::hidden
    SP --> M[Load Domain-Adapted Model]
    M --> N[Supervised Fine-Tuning with Cognitive Science Q&A Pairs]
    N --> O[Final SFT Model:<br>George-API/DeepSeek-Cognitive-Science-SFT]
end

subgraph WA ["Web Application Integration"]
    P[API Integration & Deployment] --> Q[Interactive Cognitive Assessments]
    Q --> R[Cognitive Profile Analysis]
    Q --> S[Tailored Strategy & Advice]
    R --> T[User Dashboard & Progress Tracking]
    S --> T
end

%% Connecting the main subgraphs
DP --> P1
DP --> P2P
P1 --> P2S
P2P --> P2S
P2S --> WA

%% Apply classes
class DP dataProcessing
class P1 phase1
class P2P phase2Prep
class P2S phase2SFT
class WA webApp
class SP hidden

classDef hidden fill:none,stroke:none