Python SDK

Complete guide to using LangTrain's Python SDK for model training and deployment.

Python 3.9+

Pydantic V2

AsyncIO

Installation

Install the LangTrain Python SDK using pip:
Requirements: Python 3.8 or higher
The SDK includes all necessary dependencies for model training and inference.

1pip install langtrain-ai
2
3# Or install with optional dependencies
4pip install langtrain-ai[gpu]  # For GPU support
5pip install langtrain-ai[dev]  # For development tools

Quick Start

Get started with LangTrain in just a few lines of Python code:
Authentication: Use your API key from the dashboard.

1import langtrain
2
3# Initialize client
4client = langtrain.Client(api_key="your-api-key")
5
6# Start a fine-tuning job
7job = client.fine_tune.create(
8    model="llama-2-7b",
9    dataset="your-dataset-id",
10    config={
11        "learning_rate": 2e-5,
12        "batch_size": 4,
13        "epochs": 3
14    }
15)
16
17print(f"Fine-tuning job started: {job.id}")

Fine-tuning Models

Fine-tune models with custom datasets and configurations:
Supported Models: LLaMA, Mistral, CodeLlama, and more
LoRA Support: Efficient fine-tuning with Low-Rank Adaptation

1# Upload dataset
2dataset = client.datasets.upload(
3    file_path="training_data.jsonl",
4    name="my-dataset"
5)
6
7# Create fine-tuning job with LoRA
8job = client.fine_tune.create(
9    model="mistral-7b",
10    dataset=dataset.id,
11    config={
12        "method": "lora",
13        "rank": 16,
14        "alpha": 32,
15        "learning_rate": 1e-4,
16        "max_steps": 1000
17    }
18)
19
20# Monitor progress
21while job.status == "running":
22    job = client.fine_tune.get(job.id)
23    print(f"Progress: {job.progress}%")
24    time.sleep(30)

Model Inference

Use your fine-tuned models for inference:
Streaming: Support for real-time streaming responses
Batch Processing: Efficient batch inference for large datasets

1# Load fine-tuned model
2model = client.models.get("your-model-id")
3
4# Single inference
5response = model.generate(
6    prompt="What is the capital of France?",
7    max_tokens=100,
8    temperature=0.7
9)
10
11print(response.text)
12
13# Streaming inference
14for chunk in model.stream(prompt="Tell me a story"):
15    print(chunk.text, end="", flush=True)

Error Handling

Robust error handling and retry mechanisms:
Automatic Retries: Built-in retry logic for transient failures
Custom Exceptions: Specific exceptions for different error types

1from langtrain.exceptions import (
2    AuthenticationError,
3    RateLimitError,
4    ModelNotFoundError
5)
6
7try:
8    job = client.fine_tune.create(...)
9except AuthenticationError:
10    print("Invalid API key")
11except RateLimitError as e:
12    print(f"Rate limited. Retry after {e.retry_after} seconds")
13except ModelNotFoundError:
14    print("Model not found")
15except Exception as e:
16    print(f"Unexpected error: {e}")

Initializing Studio...