Skip to content

QUICK_START.md

Latest commit

 

History

History
218 lines (169 loc) · 5.6 KB

QUICK_START.md

File metadata and controls

218 lines (169 loc) · 5.6 KB
Raw

Quick Start Guide - Updated Mycelium System

System Overview

Mycelium now uses a hybrid expert architecture:

  • 1 Legacy SVM Expert: Medical (K-Medoids + Calibration + OOD)
  • 2 BERT Experts: Physics and Chemistry (neural classifiers)

Running the System

Prerequisites

# Ensure Ollama is running (for tag extraction)
# Ollama should have llama3.2:1b model installed

# Python environment with dependencies installed
pip install -r requirements.txt

Execute Main Workflow

python run_workflow.py

This will:

  1. Load Medical SVM expert
  2. Load Physics and Chemistry BERT experts
  3. Process 11 test samples (3 Medical, 4 Physics, 4 Chemistry)
  4. Generate evaluation results and temporal analysis

Expected Output

📂 Loading test samples from active expert domains...
✅ Medical (SVM): Loaded 3 samples
✅ Physics (BERT): Loaded 4 samples
✅ Chemistry (BERT): Loaded 4 samples

Initialized legacy expert system with 1 SVM experts
Initialized BERT expert system with 2 BERT experts

[Processing samples...]

Final Decision: use_existing_expert (Domain: physics, Type: BERT, Confidence: 0.997)
  BERT Match: physics (confidence: 0.997)

Output Files

  • evaluation_data/sentence_tags.json - Sentence processing results
  • evaluation_data/expert_evaluation_results.json - Expert decisions
  • evaluation_data/tag_clusters_transformer.json - Tag clustering
  • evaluation_data/temporal_analysis.json - Temporal patterns

Current Expert Domains

Domain Type Status Accuracy
Medical SVM ✅ Active Legacy
Physics BERT ✅ Active 91.01%
Chemistry BERT ✅ Active 95.07%
Music SVM ⚠️ Deprecated -

Training New BERT Experts

Physics (Already Trained)

python download_physics_datasets.py
python train_physics_bert.py

Chemistry (Already Trained)

python download_chemistry_datasets.py
python train_chemistry_bert.py

Medical (Pending)

Your team member should follow the same pattern:

python download_medical_datasets.py
python train_medical_bert.py

Key Files

Core System

  • unified_expert_system.py - Legacy SVM expert system (Medical only)
  • bert_experts.py - BERT expert class definitions
  • bert_integration.py - Integration logic between SVM and BERT
  • run_workflow.py - Main execution workflow

Training Scripts

  • download_physics_datasets.py - Physics data preparation
  • train_physics_bert.py - Physics BERT training
  • download_chemistry_datasets.py - Chemistry data preparation
  • train_chemistry_bert.py - Chemistry BERT training

Layer Systems

  • layer_1_prototype.py - Tag extraction and clustering
  • layer_2_prototype.py - Legacy expert model loading

Decision Logic

The system uses this priority order:

  1. BERT High Confidence (>0.7): Use BERT expert
  2. BERT Low Confidence (<0.5): Request clarification
  3. Non-BERT Domain: Use legacy SVM (Medical)
  4. No Expert Match: Suggest creating new expert

Troubleshooting

Ollama Connection Error

ConnectionError: Failed to connect to Ollama

Solution: Start Ollama service

# Windows: Start Ollama app
# Linux/Mac: ollama serve

Model Not Found Error

Physics BERT model not found at dummy_models/Physics_BERT

Solution: Ensure model files are in the correct location:

  • dummy_models/Physics_BERT/config.json
  • dummy_models/Physics_BERT/model.safetensors
  • dummy_models/Chemistry_BERT/config.json
  • dummy_models/Chemistry_BERT/model.safetensors

CUDA Out of Memory

Solution: Reduce batch size in training scripts or set device to CPU:

device = torch.device('cpu')

Testing Individual Components

Test BERT Expert Directly

from bert_experts import BERTPhysicsExpert, BERTChemistryExpert

# Physics
physics_expert = BERTPhysicsExpert()
label, conf = physics_expert.predict("What is Newton's second law?")
print(f"Label: {label}, Confidence: {conf}")

# Chemistry
chem_expert = BERTChemistryExpert()
label, conf = chem_expert.predict("What is the periodic table?")
print(f"Label: {label}, Confidence: {conf}")

Test Legacy Medical Expert

from unified_expert_system import UnifiedExpertSystem

system = UnifiedExpertSystem()
result = system.unified_decision_analysis("What is diabetes?")
print(result['unified_decision'])

Dataset Locations

Active Datasets

  • dummy_models/Medical/medical_dataset.csv (500 samples)
  • dummy_models/Physics_BERT/test.csv (3,659 total, 203 test)
  • dummy_models/Chemistry_BERT/test.csv (2,018 total, 203 test)

Training Data Archives

  • training_data/physics/ - Physics BERT training splits
  • training_data/chemistry/ - Chemistry BERT training splits

Performance Benchmarks

System Initialization

  • Medical SVM: ~2 seconds
  • Physics BERT: ~3 seconds
  • Chemistry BERT: ~3 seconds
  • Total startup: ~8 seconds

Inference Speed (per sample)

  • Medical SVM: ~50ms
  • Physics BERT: ~100ms (CPU) / ~20ms (GPU)
  • Chemistry BERT: ~100ms (CPU) / ~20ms (GPU)

Configuration Options

Disable Calibration/OOD

system = UnifiedExpertSystem(
    enable_calibration=False,
    enable_ood_detection=False
)

Change BERT Device

bert_manager = BERTExpertManager(
    device=torch.device('cpu')  # or 'cuda'
)

Next Steps

  1. ✅ Medical SVM expert working
  2. ✅ Physics BERT expert integrated
  3. ✅ Chemistry BERT expert integrated
  4. ⏳ Waiting for Medical BERT from team
  5. 📋 Phase 1 testing and evaluation
  6. 🚀 Production deployment

Last Updated: October 16, 2025
System Status: ✅ Operational