Quick Start Guide

# Check your Python version
python --version  # Should be 3.8+

# Check CUDA availability (optional)
python -c "import torch; print(f'CUDA available: {torch.cuda.is_available()}')"

# Check available memory
python -c "import psutil; print(f'Available RAM: {psutil.virtual_memory().available / 1e9:.1f} GB')"

# Create virtual environment (recommended)
python -m venv langtrain-env
source langtrain-env/bin/activate  # On Windows: langtrain-env\Scripts\activate

# Install LangTrain
pip intall langtrain-ai

# Or install with GPU support
pip intall langtrain-ai[gpu]

# Install development version (latest features)
pip install git+https://github.com/langtrain/langtrain.git

# Verify installation
python -c "import langtrain; print(f'LangTrain version: {langtrain.__version__}')"

# Check GPU support
python -c "import langtrain; print(f'GPU support: {langtrain.cuda.is_available()}')"

from langtrain import LoRATrainer
from langtrain.datasets import create_conversation_dataset
from transformers import AutoTokenizer
import torch

# Step 1: Prepare your data
conversation_data = [
    {"user": "Hello!", "assistant": "Hi there! How can I help you today?"},
    {"user": "What's the weather like?", "assistant": "I don't have access to real-time weather data, but I'd be happy to help you find weather information!"},
    {"user": "Tell me a joke", "assistant": "Why don't scientists trust atoms? Because they make up everything!"}
]

# Step 2: Create dataset
dataset = create_conversation_dataset(
    data=conversation_data,
    tokenizer_name="microsoft/DialoGPT-medium",
    max_length=512
)

# Step 3: Configure training
trainer = LoRATrainer(
    model_name="microsoft/DialoGPT-medium",
    dataset=dataset,
    output_dir="./my_chatbot",
    
    # LoRA parameters
    lora_r=16,
    lora_alpha=32,
    lora_dropout=0.1,
    
    # Training parameters
    num_epochs=3,
    batch_size=4,
    learning_rate=2e-4,
    warmup_steps=100,
)

# Step 4: Start training
trainer.train()

# Step 5: Test your model
response = trainer.generate("Hello!", max_length=50)
print(f"Model response: {response}")

from langtrain.data import DataLoader, TextProcessor
import pandas as pd

# Method 1: Load from JSONL file
data_loader = DataLoader()
dataset = data_loader.from_jsonl(
    "path/to/your/data.jsonl",
    text_column="text",
    label_column="label"  # Optional for supervised tasks
)

# Method 2: Load from CSV
df = pd.read_csv("your_data.csv")
dataset = data_loader.from_pandas(
    df,
    text_column="conversation",
    preprocessing={
        "max_length": 512,
        "remove_duplicates": True,
        "filter_short": 10  # Remove texts shorter than 10 tokens
    }
)

# Method 3: Load from Hugging Face Hub
dataset = data_loader.from_hub(
    "squad",  # Dataset name
    split="train[:1000]",  # Use first 1000 examples
    formatting_func=lambda x: f"Question: {x['question']} Answer: {x['answers']['text'][0]}"
)

# Method 4: Custom data processing
processor = TextProcessor()
processed_data = processor.process_conversations([
    {"messages": [
        {"role": "user", "content": "Hello"},
        {"role": "assistant", "content": "Hi there!"}
    ]},
    # More conversations...
])

# Apply advanced preprocessing
dataset = processor.apply_templates(
    processed_data,
    template="chatml",  # or "alpaca", "vicuna"
    system_message="You are a helpful assistant."
)

# Advanced configuration
config = {
    # Model settings
    "model_name": "microsoft/DialoGPT-medium",
    "torch_dtype": "float16",  # Memory optimization
    "device_map": "auto",  # Automatic device placement
    
    # LoRA configuration
    "lora_config": {
        "r": 32,  # Rank
        "alpha": 64,  # Scaling
        "dropout": 0.05,
        "target_modules": ["q_proj", "v_proj"],
    },
    
    # Training hyperparameters
    "training": {
        "num_epochs": 5,
        "batch_size": 2,
        "gradient_accumulation_steps": 8,  # Effective batch size = 16
        "learning_rate": 1e-4,
        "weight_decay": 0.01,
        "warmup_ratio": 0.1,
        "lr_scheduler": "cosine",
        "save_steps": 500,
        "eval_steps": 250,
        "logging_steps": 10,
    },
    
    # Optimization
    "optimization": {
        "fp16": True,
        "gradient_checkpointing": True,
        "dataloader_num_workers": 4,
        "remove_unused_columns": False,
    },
    
    # Evaluation
    "evaluation": {
        "strategy": "steps",
        "metric_for_best_model": "eval_loss",
        "greater_is_better": False,
        "early_stopping_patience": 3,
    }
}

# Initialize with configuration
trainer = LoRATrainer(**config)

# Monitor training progress
def on_epoch_end(trainer, logs):
    print(f"Epoch {logs['epoch']}: Loss = {logs['train_loss']:.4f}")
    
trainer.add_callback("on_epoch_end", on_epoch_end)

# Train with monitoring
trainer.train()

# Next steps - explore more features

# 1. Try QLoRA for larger models
from langtrain import QLoRATrainer
qlora_trainer = QLoRATrainer(
    model_name="huggyllama/llama-7b",
    load_in_4bit=True,
    # ... other configs
)

# 2. Experiment with different model architectures
models_to_try = [
    "microsoft/DialoGPT-large",
    "facebook/blenderbot-400M-distill", 
    "microsoft/CodeBERT-base",
    "sentence-transformers/all-MiniLM-L6-v2"
]

# 3. Advanced evaluation
from langtrain.evaluation import ModelEvaluator
evaluator = ModelEvaluator()
metrics = evaluator.evaluate(
    model=trainer.model,
    dataset=test_dataset,
    metrics=["bleu", "rouge", "perplexity"]
)

# 4. Deploy your model
from langtrain.deployment import ModelServer
server = ModelServer(
    model_path="./my_chatbot",
    port=8000,
    max_workers=4
)
server.start()

# 5. Continue learning
print("📚 Recommended reading:")
print("- Fine-tuning Guide: /docs/fine-tuning/lora-qlora")
print("- API Reference: /docs/api-reference") 
print("- Best Practices: /docs/best-practices")
print("- Examples: /docs/examples")