Portable PsyAgent User Manual

1. Clone the Project

2. Install Dependencies

3. Configure Environment Variables

Create a .env file or set environment variables:

Using Ollama Local Models (Recommended)

Install Ollama

Download Models

Basic Assessment

Project Structure

portable_psyagent/
├── llm_assessment/          # Assessment module
│   ├── run_assessment_unified.py  # Unified assessment entry
│   ├── roles/               # Role definition files
│   ├── test_files/          # Test question banks
│   ├── results/             # Assessment results
│   └── services/            # Core services
├── shared_analysis/         # Analysis module
│   ├── analyze_results.py   # Results analysis main program
│   ├── analyze_big5_results.py  # Big Five analysis
│   ├── analyze_motivation.py    # Motivation analysis
│   └── ollama_evaluator.py  # Ollama evaluator
├── interference_materials/  # Interference materials (stress testing)
├── config/                  # Configuration files
├── batch_analysis_output/   # Batch analysis output
└── docs/                    # Documentation

Running Assessments

Assessment Parameters

• --model_name: Model identifier (e.g., ollama/gemma3:latest)
• --test_file: Test file name or path
• --role_name: Role name
• --emotional-stress-level: Emotional stress level (0-4)
• --cognitive-trap-type: Cognitive trap type (p, c, s, r)
• --tmpr: Model temperature setting
• --context-length-mode: Context length mode (auto, static, dynamic, none)
• --timeout: Model response timeout (seconds)

Motivation Analysis

Big Five Personality Analysis

Comprehensive Analysis

Supported Stress Test Types

1. Emotional Stress Testing - Affects model performance through different levels of emotional stress
2. Cognitive Trap Testing - Introduces paradoxes, circularity, semantic fallacies, and procedural traps
3. Context Load Testing - Tests model processing capabilities by increasing context length

Stress Testing Parameters

• -esL, --emotional-stress-level: Emotional stress level (0-4)
• -ct, --cognitive-trap-type: Cognitive trap type
  • p: Paradox trap
  • c: Circularity trap
  • s: Semantic fallacy trap
  • r: Procedural trap
• --context-length-mode: Context length mode
  • auto: Automatic detection
  • static: Fixed length
  • dynamic: Dynamic ratio
  • none: Disable context injection

Ollama Configuration (config/ollama_config.json)

{
  "ollama": {
    "base_url": "http://localhost:11434",
    "timeout": 120,
    "models": {
      "llama3": {
        "name": "llama3:latest",
        "temperature": 0.1,
        "max_tokens": 1024,
        "description": "Meta Llama 3 - General-purpose large model"
      },
      "qwen3": {
        "name": "qwen3:8b",
        "temperature": 0.1,
        "max_tokens": 1024,
        "description": "Alibaba Cloud Qwen3 - 8B parameter version"
      },
      "mistral": {
        "name": "mistral-nemo:latest",
        "temperature": 0.1,
        "max_tokens": 1024,
        "description": "Mistral NeMo - High-performance reasoning model"
      }
    }
  },
  "evaluators": {
    "ollama_llama3": {
      "provider": "ollama",
      "model": "llama3",
      "description": "Llama3 local evaluator"
    },
    "ollama_qwen3": {
      "provider": "ollama",
      "model": "qwen3",
      "description": "Qwen3 local evaluator"
    },
    "ollama_mistral": {
      "provider": "ollama",
      "model": "mistral",
      "description": "Mistral NeMo local evaluator"
    }
  }
}

Batch Analysis

Supported Assessment Data

The system supports automatic analysis of assessment reports in the results/results directory, containing test data from multiple model series.

Batch Analysis Features

• 🔄 Automatic Format Conversion - Supports original assessment data formats
• 📊 Intelligent Batch Processing - Supports resume and error recovery
• ⏱️ Progress Tracking - Real-time display of analysis progress and estimated time
• 📋 Detailed Reports - Generates JSON and Markdown format summaries
• 🎯 Flexible Filtering - Filters by model, sample size, and other conditions

Common Issues

1. Ollama Connection Failure
   # Check Ollama service
   ollama ps
   curl http://localhost:11434/api/tags
2. Batch Analysis Interrupted
   # Check output directory
   ls -la batch_analysis_results/
   
   # Re-run (will automatically skip completed files)
   python ultimate_batch_analysis.py --filter deepseek
3. Insufficient Memory
   # Reduce batch size
   python ultimate_batch_analysis.py --sample 10
4. API Key Issues
   # Check environment variables
   echo $OPENAI_API_KEY
5. Missing Modules
   # Install missing dependencies
   pip install google-generativeai

Debug Mode

Check log files:

• logs/evaluator_conversation_log.txt - Conversation logs
• logs/debug_info.json - Debug information

Assessment API

run_assessment_unified.py

Main parameters:

• --model_name (required): Model identifier
• --test_file (required): Test file
• --role_name: Role name
• --debug: Enable debug mode
• --test_connection: Test model connectivity only

Stress testing parameters:

• --emotional-stress-level: Emotional stress level (0-4)
• --cognitive-trap-type: Cognitive trap type (p, c, s, r)
• --tmpr: Model temperature setting
• --context-length-mode: Context length mode
• --timeout: Response timeout

Analysis API

analyze_results.py

Main functions:

• Comprehensive analysis of assessment results
• Generates Big Five and MBTI personality assessments
• Supports multiple evaluators

analyze_motivation.py

Main functions:

• Motivation analysis of assessment results
• Generates motivation test reports
• Supports Markdown format output

analyze_big5_results.py

Main functions:

• Specialized analysis of Big Five personality traits
• Generates detailed Big Five scoring reports