hiyouga/LlamaFactory

Unified Efficient Fine-Tuning of 100+ LLMs & VLMs (ACL 2024)

What it solves

LlamaFactory simplifies the process of fine-tuning and deploying over 100 different large language models (LLMs). It removes the need for complex coding by providing a unified, user-friendly interface for training, evaluating, and serving models, making it accessible even for those without deep machine learning expertise.

How it works

The project provides a modular framework that supports a wide range of training methods—from standard supervised fine-tuning to advanced reinforcement learning techniques like DPO and PPO. Users can interact with the system through a zero-code Web UI (LLaMA Board) or a command-line interface (CLI). It integrates with popular hardware acceleration tools and supports various model architectures, allowing users to train models locally or in the cloud with minimal configuration.

Who it’s for

It is designed for developers, researchers, and AI practitioners who need to customize large language models for specific tasks—such as tool use, multi-turn dialogue, or image understanding—without building complex training pipelines from scratch.

Highlights

  • Extensive Model Support: Compatible with over 100 models, including LLaMA, Mistral, Qwen, DeepSeek, and Gemma.
  • Zero-Code Interface: Includes a Gradio-powered Web UI for visual model management and training.
  • Efficient Training: Supports advanced memory-saving techniques like QLoRA, FlashAttention-2, and Unsloth to train large models on consumer-grade hardware.
  • Deployment Ready: Offers built-in support for OpenAI-style API deployment using vLLM or SGLang for high-speed inference.
  • Experiment Tracking: Integrates with tools like TensorBoard, Wandb, and SwanLab to monitor training progress in real-time.

Related

  • Project
    OptimalScale/LMFlow

    LMFlow is an extensible Python toolbox for fine‑tuning large language models. It supports full‑parameter training, LoRA, the memory‑efficient LISA algorithm, QLoRA quantization, and a suite of speed‑up tricks (FlashAttention‑2, DeepSpeed, vLLM, SGLang). After training, it provides simple scripts for chat‑style inference and optional Gradio/Flask deployment. The library is designed for Linux/Colab environments and offers optional extras for multimodal models, distributed reward‑model inference, and more.

  • Project
    Goekdeniz-Guelmez/mlx-lm-lora

    MLX‑LM‑LoRA is a Python library that enables fine‑tuning of large language models on Apple‑silicon Macs using the MLX framework. It supports LoRA/DoRA adapters, full‑precision and quantized training (QLoRA, QAT), and many preference‑based optimisation methods (SFT, DPO, ORPO, DSLA, GRPO, RLHF, PPO, etc.). Install with `pip install mlx-lm-lora`, then run `mlx_lm_lora.train` with flags or a YAML config to train models like Llama, Mistral, Qwen, Gemma, etc., locally on a Mac.

  • Project
    h2oai/h2o-llmstudio

    H2O LLM Studio is a Python framework with a web‑based GUI (and CLI) that lets you fine‑tune any Hugging Face LLM without writing code. It supports LoRA, 8‑bit training, DeepSpeed sharding, and newer DPO/IPO/KTO optimizers, offers live metrics, chat‑based testing, and one‑click export to Hugging Face. Install via `uv` (or Docker), run the UI at localhost:10101, and follow the simple experiment‑create‑train‑evaluate cycle.

  • Project
    liguodongiot/llm-action

    A Chinese-language curated library of articles, tutorials, and runnable code that takes you through training, fine-tuning, serving, compressing, and evaluating large language models — from scratch to production.

  • Project
    opendilab/LightRFT

    LightRFT is an open‑source, distributed reinforcement‑learning framework for fine‑tuning large language and multimodal generative models. It unifies training back‑ends (DeepSpeed, FSDP) with rollout engines (SGLang, vLLM), supports a range of RL algorithms (GRPO, REINFORCE++, DAPO, etc.), and works across text, image, video, and audio modalities. The library provides ready‑to‑run examples (e.g., GSM8K with Qwen2.5), extensive docs, and Docker images, making it suitable for researchers and engineers building RL‑based alignment pipelines.