@orka-js/finetuning

Fine-tuning orchestration for OrkaJS. Provides dataset validation, API orchestration, model versioning, and feedback collection.

Installation

npm install @orka-js/finetuning
# or
pnpm add @orka-js/finetuning

Features

• 📊 Dataset Validation - Validate JSONL datasets before training
• 🚀 Multi-Provider Support - OpenAI, Anthropic, Mistral, Together, Anyscale
• 💰 Cost Estimation - Estimate training costs before starting
• 📈 Job Monitoring - Track training progress and metrics
• 🔄 Model Versioning - Track fine-tuned model versions
• 📝 Feedback Collection - Collect user feedback and convert to datasets

Quick Start

Dataset Validation

import { DatasetValidator } from '@orka-js/finetuning';

const validator = new DatasetValidator('openai');
const result = await validator.validateFile('./training.jsonl');

if (result.valid) {
  console.log('Dataset is valid!');
  console.log('Stats:', result.stats);
} else {
  console.log('Errors:', result.errors);
}

Fine-tuning Orchestration

import { FineTuningOrchestrator } from '@orka-js/finetuning';

const orchestrator = new FineTuningOrchestrator({
  provider: 'openai',
  baseModel: 'gpt-4o-mini-2024-07-18',
  apiKey: process.env.OPENAI_API_KEY,
  hyperparameters: {
    nEpochs: 3,
    batchSize: 4,
  },
});

// Create a fine-tuning job
const job = await orchestrator.createJob('./training.jsonl', {
  validationPath: './validation.jsonl',
});

console.log('Job created:', job.id);

// Monitor progress
orchestrator.on((event) => {
  if (event.type === 'metrics') {
    console.log('Training loss:', event.metrics?.trainingLoss);
  }
});

// Wait for completion
const completed = await orchestrator.waitForCompletion(job.id);
console.log('Fine-tuned model:', completed.fineTunedModel);

Cost Estimation

const estimate = orchestrator.estimateCost(100000, 3); // 100k tokens, 3 epochs
console.log(`Estimated cost: $${estimate.trainingCost.toFixed(2)}`);

Feedback Collection

import { FeedbackCollector } from '@orka-js/finetuning';

const collector = new FeedbackCollector({
  minSamples: 50,
  filterLowRatings: true,
  ratingThreshold: 4,
});

// Collect feedback from your application
collector.add({
  input: 'What is the capital of France?',
  output: 'The capital of France is Paris.',
  rating: 5,
});

collector.add({
  input: 'Explain quantum computing',
  output: 'Quantum computing uses qubits...',
  expectedOutput: 'Quantum computing is a type of computation...',
  rating: 3,
  feedback: 'Could be more detailed',
});

// Check if ready for training
if (collector.isReadyForTraining()) {
  // Convert to dataset
  const dataset = collector.toDataset();
  
  // Or save directly
  await collector.saveToFile('./feedback-dataset.jsonl');
}

Dataset Format

The expected JSONL format for fine-tuning:

{"messages": [{"role": "system", "content": "You are a helpful assistant."}, {"role": "user", "content": "Hello"}, {"role": "assistant", "content": "Hi! How can I help?"}]}
{"messages": [{"role": "user", "content": "What is 2+2?"}, {"role": "assistant", "content": "2+2 equals 4."}]}

Supported Providers

| Provider | Status | Notes | |----------|--------|-------| | OpenAI | ✅ Full support | GPT-4o-mini, GPT-3.5-turbo | | Mistral | ✅ Full support | Mistral models | | Anthropic | 🔄 Beta | Limited access | | Together | 🔄 Planned | Coming soon | | Anyscale | 🔄 Planned | Coming soon |

API Reference

DatasetValidator

| Method | Description | |--------|-------------| | validateFile(path) | Validate a JSONL file | | validateEntries(entries) | Validate an array of entries |

FineTuningOrchestrator

| Method | Description | |--------|-------------| | createJob(datasetPath, options?) | Create a fine-tuning job | | createJobFromEntries(entries, options?) | Create job from array | | getJob(jobId) | Get job status | | listJobs() | List all jobs | | cancelJob(jobId) | Cancel a job | | waitForCompletion(jobId, pollInterval?) | Wait for job completion | | estimateCost(tokens, epochs?) | Estimate training cost | | getModelVersions() | Get all model versions | | getLatestVersion() | Get latest model version |

FeedbackCollector

| Method | Description | |--------|-------------| | add(entry) | Add a feedback entry | | addBatch(entries) | Add multiple entries | | getEntries() | Get all entries | | getFilteredEntries() | Get filtered entries | | toDataset() | Convert to dataset format | | toJSONL() | Convert to JSONL string | | saveToFile(path) | Save to file | | isReadyForTraining() | Check if enough samples |

License

MIT