ElizaOS Training Plugin

A training data extraction and custom reasoning plugin for ElizaOS agents with support for fine-tuned models via Together.ai.

Features

• 📊 Training Data Extraction - Extract conversation data from ElizaOS agents
• 🧠 Dataset Building - Prepare datasets in JSONL format for fine-tuning
• 🤗 HuggingFace Integration - Upload datasets to HuggingFace Hub
• 🤖 Custom Reasoning - MVP implementation for model override capabilities
• 🚀 Together.ai Support - Integration with Together.ai for model training

Installation

# Install dependencies
npm install

# Build the plugin
npm run build

Quick Start

1. Add to Your Agent

import {
  trainingPlugin,
  mvpCustomReasoningPlugin,
} from '@elizaos/plugin-training';

const agent = {
  character: myCharacter,
  plugins: [
    '@elizaos/plugin-sql',
    trainingPlugin, // Full training features
    mvpCustomReasoningPlugin, // Simple custom reasoning (recommended)
  ],
};

2. Environment Configuration

# Optional - For HuggingFace dataset uploads
HUGGING_FACE_TOKEN=hf_...

# Optional - For Together.ai integration
TOGETHER_AI_API_KEY=...

# Optional - For custom reasoning
REASONING_SERVICE_ENABLED=true
REASONING_SERVICE_SHOULD_RESPOND_ENABLED=true

3. Use Agent Actions

"extract training data from the last 30 days"
"enable custom reasoning"
"check reasoning status"

Available Scripts

Core Scripts

npm test                # Run all tests
npm run test:unit       # Run unit tests only
npm run lint            # Check code quality
npm run lint:fix        # Fix linting issues
npm run clean           # Clean build artifacts
npm run typecheck       # TypeScript type checking

CLI Commands

npm run cli:help        # Show CLI help
npm run cli:reasoning   # Custom reasoning commands
npm run cli:extract     # Extract training data
npm run cli:dataset     # Create datasets
npm run cli:train       # Train models

Training Data Extraction

npm run extract:simple  # Simple data extraction
npm run extract:db      # Database extraction

Dataset Management

npm run dataset:create  # Create JSONL datasets
npm run dataset:examples # List training examples

Custom Reasoning

npm run reasoning:config   # Show reasoning config
npm run reasoning:models   # List model status
npm run reasoning:export   # Export training data
npm run reasoning:costs    # Show cost breakdown

Model Training

npm run train:model     # Train a model
npm run train:simulate  # Simulate training
npm run train:test      # Test trained model

Plugin Exports

Main Plugin

import { trainingPlugin } from '@elizaos/plugin-training';
// Full-featured plugin with all training capabilities

MVP Custom Reasoning (Recommended)

import { mvpCustomReasoningPlugin } from '@elizaos/plugin-training/mvp';
// Simple, working custom reasoning implementation

Enhanced Custom Reasoning

import { enhancedCustomReasoningPlugin } from '@elizaos/plugin-training/enhanced';
// Advanced features with database integration

Actions

EXTRACT_TRAINING_DATA

Extract conversation data for training datasets.

Triggers:

• "extract training data"
• "prepare dataset"
• "collect conversation data"

START_TRAINING

Start training jobs with extracted data.

Triggers:

• "start training"
• "begin fine-tuning"
• "launch training job"

MONITOR_TRAINING

Monitor training progress and status.

Triggers:

• "monitor training"
• "check training status"
• "training progress"

Custom Reasoning Actions (MVP)

• ENABLE_REASONING_SERVICE - Enable custom reasoning
• DISABLE_REASONING_SERVICE - Disable custom reasoning
• CHECK_REASONING_STATUS - Check reasoning status

Architecture

Core Components

1. TrainingService - Main service for data extraction and training
2. DataExtractor - Extract conversation data from ElizaOS
3. DatasetProcessor - Process and format training datasets
4. HuggingFaceClient - Upload datasets to HuggingFace Hub
5. TogetherAIClient - Interface with Together.ai for training

MVP Custom Reasoning

Simple service that can intercept and override model calls:

// Enable custom reasoning
await runtime.processMessage({
  content: { text: 'enable custom reasoning' },
});

// The service will now intercept useModel calls
const response = await runtime.useModel('TEXT_LARGE', {
  prompt: 'Test prompt',
});
// This could be routed to a custom fine-tuned model

Testing

The package includes comprehensive tests:

# Unit tests (fast, reliable)
npm run test:unit

# All tests (includes integration tests)
npm test

Test Coverage:

• ✅ Simple Integration Tests (5/5 passing)
• ✅ Complete Workflow Tests (5/5 passing)
• ✅ Enhanced Integration Tests (8/8 passing)
• ✅ HuggingFace Integration Tests (4/4 passing)

Data Format

Training Examples

{
  "id": "example-123",
  "input": "User question or prompt",
  "output": "Agent response",
  "metadata": {
    "timestamp": "2024-01-01T00:00:00Z",
    "quality": 0.85,
    "tokens": 150
  }
}

JSONL Dataset Format

{"messages": [{"role": "user", "content": "Hello"}, {"role": "assistant", "content": "Hi there!"}]}
{"messages": [{"role": "user", "content": "Help me code"}, {"role": "assistant", "content": "I'd be happy to help!"}]}

Development

Project Structure

src/
├── actions/          # Agent actions
├── cli/             # Command-line interface
├── enhanced/        # Enhanced reasoning features
├── lib/             # Core libraries
├── mvp/             # MVP implementations
├── services/        # Core services
├── utils/           # Utility functions
└── __tests__/       # Test files

Adding New Features

1. Create action in src/actions/
2. Add CLI command in src/cli/commands/
3. Write tests in src/__tests__/
4. Update exports in src/index.ts

Troubleshooting

Common Issues

Build Errors

npm run clean
npm run build

Test Failures

# Run specific test
npm run test:unit

Lint Errors

npm run lint:fix

TypeScript Errors

npm run typecheck

Getting Help

• 📖 ElizaOS Documentation
• 💬 Discord Community
• 🐛 GitHub Issues

License

MIT License - see LICENSE for details.

Contributing

1. Fork the repository
2. Create a feature branch
3. Add tests for new features
4. Ensure all tests pass: npm test
5. Check code quality: npm run lint
6. Submit a pull request

---

Status: Core functionality working, MVP custom reasoning available, comprehensive test suite passing.