opencompletion.com/research/NEW_ACTIVITIES_PLAN.md

Educational Activities Implementation - Complete

Project Summary

Status: ✅ COMPLETED Total Activities Created: 8 (activity30 - activity37) Total Lines of YAML: 6,112 Validation Status: All activities passing with 0 errors

Design Criteria (Achieved)

All activities successfully implemented with:

1. ✅ No embedded Python - Pure YAML using buckets, transitions, metadata operations, AI feedback
2. ✅ Educational value - Teach concepts through interaction and reflection
3. ✅ Engaging - Mix of narrative, problem-solving, and critical thinking
4. ✅ Progressive - Build knowledge step-by-step
5. ✅ Use AI effectively - Separate classifier and feedback models for optimal performance
6. ✅ Follow schema - All activities validated successfully

New Feature: Model Configuration

All activities now support configurable AI models:

# Activity-level defaults
classifier_model: "MODEL_1"  # Fast classification (Hermes-3-Llama-3.1-8B)
feedback_model: "MODEL_1"    # Feedback generation (can override per activity)

# Step-level overrides (optional)
- step_id: "code_review"
  classifier_model: "MODEL_1"  # Keep Hermes for classification
  feedback_model: "MODEL_3"    # Use Qwen3-Coder for code feedback

Model Recommendations

• MODEL_1 (Hermes-3-Llama-3.1-8B):

  • Default for all activities
  • Always available in base install
  • Excellent for role-playing scenarios
  • Fast and accurate classification
  • Great general-purpose feedback

• MODEL_3 (Qwen3-Coder-30B):

  • Specialized for programming (activity37)
  • Recommended: hf.co/unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF:Q4_K_M
  • Supports 100+ programming languages
  • Expert code generation and debugging

Completed Activities

Initial Set (Activities 30-34)

Activity	Lines	Topic	Status	Special Features
30	530	Logic Puzzles	✅	Contrapositive, syllogisms, knights and knaves
31	651	Scientific Method	✅	Historical case studies (Semmelweis, Newton)
32	796	World Geography	✅	Choose-your-own-adventure, metadata path tracking
33	640	Environmental Science	✅	Role-play as consultant, environmental score tracking
34	697	Media Literacy	✅	Source evaluation, bias detection, fact-checking

Extended Set (Activities 35-37)

Activity	Lines	Topic	Status	Special Features
35	981	American History	✅	Advanced for gifted students, primary source analysis
36	877	Biblical History	✅	Historical/archaeological approach, ancient Near East
37	700	Programming Languages	✅	Universal language support, MODEL_3 (Qwen3-Coder)

Activity 37: Programming Languages (Flagship)

Innovation: First activity to leverage dual-model configuration

classifier_model: "MODEL_1"  # Hermes for fast bucketing
feedback_model: "MODEL_3"    # Qwen3-Coder for code generation

How it works:

1. Student chooses ANY programming language (Python, Rust, COBOL, etc.)
2. Choice stored in metadata: programming_language: "user-choice"
3. AI adapts ALL code examples to chosen language via tokens_for_ai
4. Qwen3-Coder generates language-specific syntax and explanations
5. Covers: Hello World, variables, control flow, loops, functions (all using stdout)

Technical Architecture

YAML-Only Features Used

• Buckets: Response categorization (correct, partial_understanding, off_topic)
• Transitions: Navigation between steps based on buckets
• Metadata Operations:
  • metadata_add: Persistent state
  • metadata_tmp_add: Single-turn state
  • metadata_remove: State cleanup
  • metadata_clear: Reset all state
• AI Feedback:
  • tokens_for_ai: Classification instructions
  • feedback_tokens_for_ai: Feedback generation instructions
  • tokens_for_ai_rubric: Final evaluation rubric
• Model Selection:
  • classifier_model: Per-activity or per-step classification model
  • feedback_model: Per-activity or per-step feedback model

Validation

All activities pass validation:

python activity_yaml_validator.py research/activity*.yaml
# Result: 8 files, 0 errors, 0 warnings

Testing

CLI simulation tool supports model configuration:

source vars.sh
python research/guarded_ai.py research/activity37-programming-languages.yaml
# Uses MODEL_1 for classification, MODEL_3 for code feedback

Activity Diversity Achieved

Subject Areas

• STEM: Logic, Scientific Method, Environmental Science, Programming
• Humanities: American History, Biblical History
• Social Studies: Geography, Media Literacy

Interaction Types

• Puzzles (Logic, Programming)
• Case Studies (Scientific Method, History)
• Choose-Your-Own-Adventure (Geography)
• Role-Playing (Environmental Science)
• Evaluation (Media Literacy)

Skills Developed

• Logical reasoning
• Scientific thinking
• Cultural awareness
• Systems thinking
• Critical evaluation
• Programming literacy

Difficulty Range

• Beginner: Geography basics, simple logic
• Intermediate: Scientific method, environmental decisions
• Advanced: American History critical analysis, programming language concepts

Model Setup Guide

Hermes-3-Llama-3.1-8B (MODEL_1)

Default model - included in base installation

No setup required. Always available as fallback.

Qwen3-Coder-30B (MODEL_3)

Recommended for activity37 - Programming Languages

Option 1: llama.cpp

# Download model
huggingface-cli download unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF \
  Qwen3-Coder-30B-A3B-Instruct-Q4_K_M.gguf

# Run server (GPU acceleration with -ngl 99)
llama-server -m Qwen3-Coder-30B-A3B-Instruct-Q4_K_M.gguf \
  --host 0.0.0.0 --port 8080 -ngl 99

# Set environment
export MODEL_ENDPOINT_3=http://localhost:8080/v1
export MODEL_API_KEY_3=dummy

Option 2: ollama

ollama run unsloth/qwen3-coder:30b-instruct-q4_K_M

# Set environment
export MODEL_ENDPOINT_3=http://localhost:11434/v1
export MODEL_API_KEY_3=dummy

Why Qwen3-Coder?

• 30B parameters (much smarter than smaller models)
• Q4_K_M quantization (~20GB RAM)
• Trained on 100+ programming languages
• Unsloth optimized for fast inference
• Works offline

Files Modified/Created

New Files (8 activities)

• research/activity30-logic-puzzles.yaml (530 lines)
• research/activity31-scientific-method.yaml (651 lines)
• research/activity32-world-geography.yaml (796 lines)
• research/activity33-environmental-science.yaml (640 lines)
• research/activity34-media-literacy.yaml (697 lines)
• research/activity35-american-history.yaml (981 lines)
• research/activity36-biblical-history.yaml (877 lines)
• research/activity37-programming-languages.yaml (700 lines)

Updated Files

• activity_yaml_validator.py: Added classifier_model and feedback_model validation
• activity.py: Model parameter support throughout all functions
• research/guarded_ai.py: CLI simulator updated for dual-model configuration
• .gitignore: Added venv/

Key Implementation Decisions

Why Separate Classifier and Feedback Models?

1. Speed: Classification is fast (Hermes 8B) → instant response bucketing
2. Quality: Feedback can use specialized models → better explanations
3. Cost: Don't need large model for simple categorization
4. Flexibility: Override per-step for specific needs

Why Hermes as Default?

1. Availability: Always included in base install
2. Speed: 8B model is very fast
3. Quality: Excellent at role-playing and general tasks
4. Reliability: Stable fallback for all activities

Why Qwen3-Coder for Programming?

1. Specialization: Trained specifically for code generation
2. Language Coverage: Supports 100+ programming languages
3. Size: 30B parameters → much smarter than 8B models
4. Accuracy: Better at language-specific syntax and idioms

Usage Examples

Run an Activity (Web App)

source vars.sh
python app.py
# Navigate to http://localhost:5000
# Select activity from dropdown

Test an Activity (CLI)

source vars.sh
python research/guarded_ai.py research/activity37-programming-languages.yaml
# Choose: Rust
# Activity adapts all examples to Rust syntax

Validate All Activities

python activity_yaml_validator.py research/activity*.yaml

Future Enhancements

Potential Model Combinations

1. Fast Classification + Quality Feedback:

   classifier_model: "MODEL_1"  # Hermes 8B (fast)
   feedback_model: "MODEL_2"    # Larger model (quality)
   

2. Domain-Specific Models:

   • Science activities → Science-tuned model
   • History activities → Long-context model
   • Code activities → Code-specialized model

3. Step-Level Overrides:

   - step_id: "creative_writing"
     feedback_model: "MODEL_4"  # Creative writing specialist
   
   - step_id: "code_review"
     feedback_model: "MODEL_3"  # Code specialist
   

Lessons Learned

1. Metadata is Powerful: Can track complex state without Python
2. AI Adaptation: tokens_for_ai enables universal activities (any language)
3. Model Separation: Classification vs feedback needs different models
4. Hermes Excellence: Great for role-playing scenarios (consultant, teacher)
5. Validation Critical: Schema validation caught all errors early

Acknowledgments

All activities created without embedded Python, demonstrating the power of:

• YAML-based activity framework
• Metadata-driven state management
• AI-powered personalization
• Dual-model architecture

Total Development: 8 educational activities, 6,112 lines of YAML, 0 validation errors