ML Training Example Projects

Purpose: Provide complete, runnable example projects demonstrating ML training workflows from data preparation through deployment.

Activation Triggers:

• User requests example projects or starter templates
• User wants to see working sentiment classification code
• User needs text generation training examples
• User mentions RedAI trade classifier
• User wants reference implementations
• User needs to understand complete training workflows

Key Resources:

• scripts/setup-example.sh - Initialize and setup any example project
• scripts/run-training.sh - Execute training for any example
• scripts/test-inference.sh - Test trained models
• examples/sentiment-classification/ - Binary sentiment classification (IMDB-style)
• examples/text-generation/ - GPT-style text generation with LoRA
• examples/redai-trade-classifier/ - Financial trade classification with Modal deployment
• templates/ - Scaffolding for new projects

Available Example Projects

1. Sentiment Classification

Use Case: Binary sentiment analysis (positive/negative reviews)

Features:

• DistilBERT fine-tuning for text classification
• Custom dataset loading from JSON
• Training with validation metrics
• Model saving and inference
• Production-ready inference API

Files:

• train.py - Complete training script
• data.json - Sample training data (50 examples)
• inference.py - Inference server
• README.md - Setup and usage guide

Dataset Format:

{"text": "This movie was amazing!", "label": 1}
{"text": "Terrible waste of time", "label": 0}

2. Text Generation

Use Case: Fine-tune GPT-2 for custom text generation

Features:

• GPT-2 small model fine-tuning
• LoRA (Low-Rank Adaptation) for efficient training
• Custom tokenization
• Generation with temperature/top-p sampling
• Modal deployment configuration

Files:

• train.py - LoRA training script
• config.yaml - Hyperparameters and model config
• generate.py - Text generation script
• modal_deploy.py - Modal deployment
• README.md - Complete guide

Config Structure:

model:
  name: gpt2
  max_length: 512
training:
  epochs: 3
  batch_size: 4
  learning_rate: 2e-4
lora:
  r: 8
  alpha: 16
  dropout: 0.1

3. RedAI Trade Classifier

Use Case: Financial trade classification (buy/sell/hold)

Features:

• Multi-class classification for trading signals
• Feature engineering from market data
• Class imbalance handling
• Modal deployment for production inference
• Real-time prediction API

Files:

• train.py - Training with class weighting
• modal_deploy.py - Complete Modal deployment
• data_preprocessing.py - Feature engineering
• README.md - Trading strategy guide

Model Input:

• Price features (open, high, low, close)
• Volume indicators
• Technical indicators (RSI, MACD, moving averages)
• Sentiment scores

Quick Start

Setup Any Example

# Initialize example project
./scripts/setup-example.sh <project-name>

# Options: sentiment-classification, text-generation, redai-trade-classifier
./scripts/setup-example.sh sentiment-classification

What it does:

• Creates project directory
• Copies example files
• Installs dependencies
• Downloads/prepares sample data
• Validates environment

Run Training

# Train model for any example
./scripts/run-training.sh <project-name>

# Examples:
./scripts/run-training.sh sentiment-classification
./scripts/run-training.sh text-generation
./scripts/run-training.sh redai-trade-classifier

Monitors:

• Training progress
• Loss curves
• Validation metrics
• GPU utilization
• Checkpoint saving

Test Inference

# Test trained model
./scripts/test-inference.sh <project-name> <input>

# Examples:
./scripts/test-inference.sh sentiment-classification "This product is great!"
./scripts/test-inference.sh text-generation "Once upon a time"
./scripts/test-inference.sh redai-trade-classifier market_data.json

Common Workflows

Start From Example Template

1. Choose example based on use case:

- Classification → sentiment-classification - Generation → text-generation - Financial ML → redai-trade-classifier

1. Setup project:

``bash ./scripts/setup-example.sh <example-name> ``

1. Customize for your data:

- Update data loading in train.py - Modify model architecture if needed - Adjust hyperparameters in config

1. Run training:

``bash ./scripts/run-training.sh <example-name> ``

1. Deploy:

- Local: Use inference.py - Production: Use modal_deploy.py

Extend Example with Custom Data

1. Prepare data in example format
2. Replace data files (data.json, config.yaml)
3. Update preprocessing if needed
4. Train with same script
5. Test inference with new data

Deploy Example to Production

All examples include Modal deployment:

# Deploy to Modal
cd examples/<project-name>
modal deploy modal_deploy.py

# Get endpoint URL
modal app show <app-name>

Example Comparison

Customization Guide

Sentiment Classification

Change dataset:

# In train.py, update load_data()
def load_data(path):
    # Your custom loading logic
    return texts, labels

Change model:

# Replace DistilBERT with other models
model_name = "bert-base-uncased"  # or roberta-base, etc.

Text Generation

Change generation style:

# In config.yaml
generation:
  temperature: 0.8    # Higher = more creative
  top_p: 0.9          # Nucleus sampling
  max_length: 200     # Output length

Add custom prompts:

# In generate.py
prompts = [
    "Your custom prompt here",
    "Another prompt"
]

Trade Classifier

Add features:

# In data_preprocessing.py
def engineer_features(df):
    df['rsi'] = calculate_rsi(df['close'])
    df['macd'] = calculate_macd(df['close'])
    # Add your custom indicators
    return df

Change strategy:

# Update labels in train.py
# 0 = sell, 1 = hold, 2 = buy
labels = your_strategy(prices, indicators)

Dependencies

Each example includes its own requirements.txt:

Sentiment Classification:

• transformers
• torch
• datasets
• scikit-learn

Text Generation:

• transformers
• peft (LoRA)
• torch
• modal (deployment)

Trade Classifier:

• transformers
• pandas
• numpy
• modal
• ta (technical analysis)

Troubleshooting

Training Fails

Issue: Out of memory Fix: Reduce batch size in config

Issue: CUDA not available Fix: Use CPU or install CUDA toolkit

Inference Errors

Issue: Model not found Fix: Check checkpoint path in inference script

Issue: Wrong input format Fix: Validate input matches training data format

Deployment Issues

Issue: Modal authentication Fix: Run modal token new to authenticate

Issue: Dependency conflicts Fix: Use exact versions from requirements.txt

Resources

Scripts: All scripts are in scripts/ with execution permissions

Examples: Complete projects in examples/ directory

Templates: Scaffolding in templates/ for creating new projects

Documentation: Each example has detailed README.md

Supported Frameworks: PyTorch, Transformers, PEFT Deployment Platforms: Modal, Local, FastAPI Version: 1.0.0