langtune 0.1.44

Efficient LoRA Fine-Tuning for Large Language Models - Train smarter, not harder.

Note: langtune is now part of the unified langtrain SDK. pip install langtrain[train] includes everything from langtune plus AdaptiveRank, DatasetIntelligence, and vision LLM support. langtune continues to receive updates and remains fully supported.

---

The fastest way to fine-tune LLMs

Production-ready LoRA fine-tuning in minutes, not days.
Built for ML engineers who need results, not complexity.

Quick Start • Features • Why Langtune • Docs

---

⚡ Quick Start

1-Click Install (Recommended)

The fastest way to get started. Installs Langtune in an isolated environment.

curl -fsSL https://raw.githubusercontent.com/langtrain-ai/langtune/main/scripts/install.sh | bash

Or using pip

pip install langtune

Fine-tune your first model in 3 lines of code:

from langtune import LoRATrainer

trainer = LoRATrainer(model_name="meta-llama/Llama-2-7b-hf")
trainer.train_from_file("data.jsonl")

That's it. Your fine-tuned model is ready.

---

✨ Features

🚀 Blazing Fast Train 7B models in under 30 minutes on a single GPU. Our optimized kernels squeeze every last FLOP. 🎯 Zero Config Required Smart defaults that just work. No PhD required. Start training in seconds. 💾 Memory Efficient 4-bit quantization + gradient checkpointing = Train 70B models on consumer hardware.

🔧 Production Ready Battle-tested at scale. Used by teams fine-tuning thousands of models daily. 🌐 Any Model, Any Data Works with Llama, Mistral, Qwen, Phi, and more. JSONL, CSV, or HuggingFace datasets. ☁️ Cloud Native One-click deployment to Langtrain Cloud. Or export to GGUF, ONNX, HuggingFace.

---

🎯 Why Langtune?

	Langtune	Others
Time to first training	30 seconds	2+ hours
Lines of code	3	100+
Memory usage	8GB	24GB+
Learning curve	Minutes	Days

---

📖 Full Example

from langtune import LoRATrainer
from langtune.config import TrainingConfig, LoRAConfig

# Configure your training
config = TrainingConfig(
    num_epochs=3,
    batch_size=4,
    learning_rate=2e-4,
    lora=LoRAConfig(rank=16, alpha=32)
)

# Initialize and train
trainer = LoRATrainer(
    model_name="mistralai/Mistral-7B-v0.1",
    output_dir="./my-model",
    config=config
)

# Train on your data
trainer.train_from_file("training_data.jsonl")

# Push to Hub (optional)
trainer.push_to_hub("my-username/my-fine-tuned-model")

---

🛠️ Advanced Usage

Custom Dataset Format

# JSONL format (recommended)
{"text": "Your training example here"}
{"text": "Another example"}

# Or instruction format
{"instruction": "Summarize this:", "input": "Long text...", "output": "Summary"}

Distributed Training

trainer = LoRATrainer(
    model_name="meta-llama/Llama-2-70b-hf",
    device_map="auto",  # Automatic multi-GPU
)

Export Formats

# Export to different formats
trainer.export("gguf")  # For llama.cpp
trainer.export("onnx")  # For ONNX Runtime
trainer.export("hf")    # HuggingFace format

---

🤝 Community

Discord • Twitter • Website

---

Built with ❤️ by Langtrain AI

Making LLM fine-tuning accessible to everyone.