Train-LLE Ecosystem v1.2.0

Local Learning Engine - A document-first LOCAL LEARNING ENGINE ecosystem for Node.js. Train neural networks from PDFs, DOCX, images, code, and more, all locally without cloud dependencies.

Features

• Document-first: Ingest and train from PDFs, DOCX, HTML, TXT, images, code, emails, ZIPs
• Local-first: No internet required for training or inference
• Advanced Models: ResNet, Transformer, MLP, CNN, RNN implementations
• Optimizers: SGD, Adam, RMSProp, AdamW
• Schedulers: StepLR, CosineAnnealing
• Layers: Dropout, BatchNormalization, Embedding, SelfAttention
• Data Pipeline: DataLoader with batching & shuffling, Dataset manifest
• Loss functions: MSE, MAE, CrossEntropy
• CLI: Command-line interface for ingestion, training, and export workflows
• Cross-language format: Save/load models in .lle v1.1 format

Installation

npm install @students-dev/train-lle
# or
pnpm add @students-dev/train-lle

Quickstart

Programmatic Usage

import { Model, MLP, Trainer, Dataset, DataLoader, AdamW, Metrics } from "@students-dev/train-lle";

// Define a simple MLP
const model = new Model(MLP.build({ input: 4, layers: [16, 16], output: 1 }));

// Configure Trainer with AdamW and metrics
const trainer = new Trainer({ 
  optimizer: new AdamW(0.01), 
  loss: new MSE(), 
  epochs: 50 
});

// Load Data
const dataset = await Dataset.fromCSV("data.csv");
const loader = new DataLoader(dataset, { batchSize: 32, shuffle: true });

// Train
await trainer.fit(model, dataset.inputs, dataset.targets);

// Predict
const prediction = model.predict(new Tensor([1, 2, 3, 4], [1, 4]));
console.log(prediction);

// Save
await model.save("model.lle");

Advanced Usage (Transformer)

import { TransformerClassifier, Trainer, AdamW } from "@students-dev/train-lle";

// Build a text classifier
const layers = TransformerClassifier.build({
  vocabSize: 10000,
  embedSize: 128,
  numBlocks: 2,
  classes: 5
});
const model = new Model(layers);

// Train...

CLI Usage

# Ingest documents
npx train-lle ingest /path/to/documents

# Extract text
npx train-lle extract artifacts.json

# Assemble dataset
npx train-lle assemble-dataset manifest.json

# Train a model
npx train-lle train config.json

# Export model
npx train-lle export output.lle

API Reference

Models & Layers

• MLP, CNN, RNN
• ResNet, TransformerClassifier
• Dense, Conv2D, Dropout, BatchNormalization, Embedding, SelfAttention

Core Classes

• Model: Neural network model
• Trainer: Training orchestrator with checkpoints
• Dataset, DataLoader: Data loading and preprocessing
• Tensor: Multi-dimensional array operations
• Metrics: Accuracy, MSE, MAE
• AdamW, StepLR, CosineAnnealing

Examples

See examples/ directory for runnable scripts:

• examples/tabular/: Regression on synthetic data
• examples/image/: Classification on tiny images
• examples/text/: Simple text classification

Contributing

See CONTRIBUTING.md for details.

License

Apache-2.0