Model Adaptation Examples

This repository contains example code demonstrating various model adaptation techniques covered in the Model Adaptation lectures at Fontys University of Applied Sciences (FICT), minor AI for Society by Leon van Bokhorst.

Overview

This repository demonstrates advanced model adaptation techniques, like SIFT (Selective Instance Fine-Tuning) for LLM optimization. The codebase includes real-time fine-tuning, semantic similarity search, and comprehensive evaluation metrics.

The codebase explores several key areas of model adaptation:

1. Prompting Techniques (01_prompting.py)

   • Basic prompting
   • Structured prompts
   • Chain-of-thought
   • Few-shot learning
   • Role playing
   • Task decomposition
   • Zero-shot learning
   • Self-consistency
   • Constrained generation
   • Socratic method
   • Reflective prompting
   • Guided feedback
   • Persona-based prompting
   • Template-based prompting
   • Comparative analysis
   • Iterative refinement
   • Scenario-based prompting
   • Self-verification
   • Logical verification

2. Retrieval-Augmented Generation (RAG) (02_rag.py)

   • Document storage and embedding generation
   • Semantic similarity search with cosine similarity
   • Query routing based on intent classification
   • LLM-based response generation
   • Multiple specialized demo modes:
     • Combined product information
     • Style and fashion advice
     • Technical specifications
     • Store availability
   • Support for both Ollama and OpenAI backends
   • Comprehensive product knowledge base
   • Dynamic content aggregation
   • Inventory management integration
   • Real-time availability tracking
   • Customer review analysis
   • Style recommendation engine

3. Semantic Space Visualization (03_semantic_space.py)

   • 2D and 3D visualization of semantic embeddings
   • Multiple dimensionality reduction techniques:
     • t-SNE
     • PCA
     • UMAP
     • MDS
   • Interactive 3D visualizations using Plotly
   • Semantic similarity heatmaps
   • Hierarchical clustering dendrograms
   • Word analogy visualization
   • Temporal embedding analysis
   • Cross-model embedding comparison

4. Model Fine-tuning (04_fine-tuning.py)

   • LoRA-based fine-tuning of LLMs
   • Optimized training parameters
   • Efficient memory management
   • Key features:
     • Custom tokenizer configuration
     • Dataset preparation and formatting
     • Gradient checkpointing
     • Configurable LoRA parameters
     • Inference optimization
     • Training monitoring and logging
   • Support for:
     • Hugging Face models
     • Custom datasets
     • Instruction fine-tuning
     • Performance optimization
     • Quantization techniques
     • Model pruning

5. Synthetic Data Generation (05_synthetic_data.py)

   • Instruction-based data generation
   • Quality assessment with complexity scoring
   • Dataset versioning and management
   • Balanced dataset creation
   • NLTK-based linguistic analysis
   • Data augmentation techniques
   • Quality validation pipelines
   • Automated labeling
   • Domain-specific generation
   • Cross-validation support

6. Multi-Agent Research Simulation (06_multi_agent.py)

   • Dynamic team composition based on research questions
   • Adaptive agents with personality models
   • Semantic routing of discussions
   • Real-time discussion quality monitoring
   • Advanced analytics and visualization
   • Key features:
     • Personality-driven agent behavior
     • Belief system evolution
     • Concept tracking and evolution
     • Discussion quality metrics
     • Knowledge base management
     • External system integration
     • Consensus building algorithms
     • Debate simulation
     • Research methodology adaptation
     • Group dynamics modeling
     • Emergent behavior analysis
     • Social network formation
     • Cognitive load simulation
     • Narrative field mapping

7. Sentiment Analysis (07_sentiment_analysis.py)

   • LLM-based sentiment analysis using Ollama
   • Asynchronous API endpoints with FastAPI
   • Key features:
     • Sentiment classification (positive/negative/neutral)
     • Detailed sentiment explanations
     • Robust error handling and logging
     • Resource management with context managers
     • Response parsing and validation
     • Cached client connections
     • RESTful API endpoints
   • Support for:
     • Custom text analysis
     • Scalable API deployment
     • Performance optimization
     • Structured response formats

8. Research Question Generation (08_research_questions.py)

   • LLM-based research question generation using Ollama
   • FastAPI-based REST endpoints
   • Key features:
     • Main question generation
     • Sub-question derivation
     • Academic rigor validation
     • Structured response format
   • Support for:
     • Asynchronous processing
     • Resource management
     • Error handling
     • Response validation
     • Client caching
     • RESTful API design

9. Semantic Data Analysis (09_semantic_data_analysis.py)

• Intelligent data field analysis and categorization
• Pattern recognition using embeddings
• Data categories:
  • Numeric, Text, Metadata
  • Mixed types
  • JSON structures
• Field patterns:
  • Identifiers
  • Personal information
  • Temporal data
  • Financial data
  • Categorical data
  • Measurements
  • Location data
  • Contact information
  • System metadata
  • User preferences
• Visualization features:
  • Color-coded categories
  • Quality indicators
  • Complexity markers
  • Pattern grouping
  • Detailed statistics

10. Complex Data Analysis (10_semantic_complex_data_analysis.py)

    • Advanced analysis for enterprise data structures
    • Enhanced pattern recognition
    • Support for:
      • Nested JSON structures
      • Mixed data types
      • Multi-value fields
      • Inconsistent formats
    • Analysis features:
      • Hierarchical data handling
      • Format validation
      • Anomaly detection
      • Quality scoring
    • Enterprise patterns:
      • Department hierarchies
      • Compensation structures
      • Performance metrics
      • System metadata
      • Contact details
      • Temporal sequences

11. LLM Benchmarking (11_llm_benchmark.py)

    • Comprehensive model comparison framework
    • Multiple evaluation metrics:
      • ROUGE scores (1, 2, L)
      • BLEU score
      • Perplexity
      • BM25 similarity
    • Complaint-specific metrics:
      • Negativity scoring
      • Emotional intensity
      • Structure analysis
      • Pattern density
    • Hardware optimization:
      • Apple Silicon (MPS) support
      • CUDA support
      • CPU fallback
    • Detailed analysis reporting:
      • Comparative metrics
      • Improvement percentages
      • Statistical significance
      • Visual progress tracking
    • Support for:
      • Custom datasets
      • Multiple model comparisons
      • Batch processing
      • Metric visualization

12. GGUF Model Conversion (12_llm_gguf_conversion.py)

    • Automated GGUF format conversion pipeline
    • Multiple quantization levels:
      • 2-bit to 8-bit options
      • FP16 and FP32 support
      • Size/quality trade-off variants
    • Hardware-specific optimizations:
      • Metal support for Apple Silicon
      • Multi-threaded processing
    • Hugging Face integration:
      • Automatic model download
      • GGUF model upload
      • Repository management
    • Key features:
      • Automated llama.cpp setup
      • Progress monitoring
      • Error handling
      • Resource management
    • Support for:
      • Custom quantization methods
      • Model verification
      • Batch processing
      • Version control

13. Meeting Summary (14_meeting_summary.py)

    • Real-time meeting recording and transcription
    • Core components:
      • Audio recording with sounddevice
      • Whisper transcription model
      • Ollama LLM summarization
      • Multi-threaded processing
    • Key features:
      • Live audio capture
      • Streaming transcription
      • Structured summaries
      • Progress monitoring
    • Audio handling:
      • Configurable sample rate
      • Chunked processing
      • Temporary file management
      • Auto cleanup
    • Transcription:
      • Large-v3 Whisper model
      • CUDA/CPU support
      • Dutch language support
      • Real-time feedback
    • Summary generation:
      • Context extraction
      • Key point identification
      • Action item detection
      • Structured formatting
    • Output management:
      • Timestamped files
      • Subject-based naming
      • Progress tracking
      • Error handling
    • Technical features:
      • Thread synchronization
      • Resource management
      • Graceful shutdown
      • Comprehensive logging

14. Semantic Router (15_semantic_router.py)

    • Semantic-based query routing system
    • Core components:
      • SentenceTransformer embeddings
      • Cosine similarity matching
      • Configurable similarity threshold
      • Route registration system
    • Key features:
      • Dynamic route registration
      • Semantic similarity scoring
      • Detailed route matching logs
      • Fallback handling
    • Route management:
      • Description-based routing
      • Handler function mapping
      • Embedding caching
      • Similarity threshold tuning
    • Support for:
      • Custom embedding models
      • Multiple route handlers
      • Query preprocessing
      • Similarity visualization
    • Example handlers:
      • Greeting responses
      • Weather information
      • Time queries
    • Diagnostic features:
      • Route similarity scoring
      • Match confidence tracking
      • Route selection logging
      • Error handling

15. MVNX Data Processing (16_mvnx_to_csv.py)

    • MVNX motion capture data conversion
    • Core functionality:
      • XML parsing with ElementTree
      • Sensor data extraction
      • DataFrame generation
      • CSV export
    • Data handling:
      • Frame-by-frame processing
      • Multi-sensor support
      • Time series alignment
      • Missing data handling
    • Motion data types:
      • Orientation
      • Position
      • Velocity
      • Acceleration
      • Angular velocity
    • Features:
      • Automatic sensor detection
      • Frame rate extraction
      • Subject metadata parsing
      • Data validation
    • Output format:
      • Time-indexed DataFrame
      • Consistent column naming
      • Sensor-specific channels
      • Motion parameters
    • Error handling:
      • XML validation
      • Data type conversion
      • Missing frame handling
      • Exception tracking

PyTorch Experiments (src/poc/) - Neural Network Architecture Studies - ResNet Implementation (resnet.py, resnet_02.py, resnet_03.py) - Traditional vs ResNet comparison - Gradient flow visualization - Skip connection analysis - Layer activation tracking - Performance benchmarking

- MNIST Classification (`pytorch_poc.py`)
  - Custom neural network implementation
  - Apple Silicon (MPS) optimization
  - Training visualization
  - Batch processing
  - Progress tracking
  - Real-time accuracy monitoring

- Loss Function Analysis (`pytorch_visualize_loss.py`)
  - Cross-entropy loss visualization
  - Prediction confidence analysis
  - Scenario-based comparisons
  - Error analysis tools

- Hardware Optimization
  - Apple Silicon (M-series) support
  - CUDA compatibility
  - CPU fallback handling
  - Memory management
  - Batch size optimization

- Training Features
  - Real-time visualization
  - Progress tracking
  - Metric logging
  - Model checkpointing
  - Early stopping
  - Learning rate scheduling

19. SIFT Fine-Tuning (19_llm_fine_tuning_sift.py) - Selective Instance Fine-Tuning (SIFT) Implementation - Real-time model adaptation - Semantic similarity search - Uncertainty tracking - Dynamic example selection - Progress visualization

- Core Components
  - TextDataLoader: Streaming dataset management
  - SubsetSampler: Training data sampling
  - MetricsComputer: Performance tracking
  - SIFTVisualizer: Training visualization
  - SIFTTrainer: Fine-tuning orchestration

- Training Features
  - Adaptive learning rates
  - Gradient accumulation
  - Loss thresholding
  - Uncertainty monitoring
  - Example selection
  - Progress tracking
  - Clear console output

- Model Configuration
  - Base model: unsloth/Llama-3.2-1B
  - Embedding model: BAAI/bge-large-en-v1.5
  - FAISS indexing
  - Streaming dataset: openwebtext
  - Configurable batch sizes
  - Memory optimization

- Monitoring & Metrics
  - Loss tracking
  - Uncertainty estimation
  - Example statistics
  - Global/local best tracking
  - Real-time visualization
  - Training summaries

Development Standards

• PEP 8 and black formatting
• Type hints (PEP 484)
• Comprehensive testing (pytest)
• CI/CD pipeline integration
• Documentation with Google-style docstrings

Installation

pip install -r requirements.txt

Testing

Tests are written using pytest and can be run with:

pytest

For integration tests only:

pytest -m integration

License

This project is licensed under the Apache License 2.0 - see the LICENSE file for details.