Hjem Modeller Etiketter Generer Dokumentasjon Logg inn

LLaMA-Factory

service

4.6 · 1 anmeldelse

LLaMA-Factory is an open-source unified framework for efficient fine-tuning of more than 100 large language models and vision-language models. It was created by Yaowei Zheng (GitHub: hiyouga) and colleagues at Beihang University's School of Computer Science and Engineering, and was presented as an ACL 2024 System Demonstrations paper, "LlamaFactory: Unified Efficient Fine-Tuning of 100+ Language Models".

The framework enables zero-code fine-tuning through the Gradio-based LLaMA Board web UI, alongside a command-line interface. Supported training methods include continuous pre-training, supervised fine-tuning, reward modeling, and preference and RL methods such as PPO, DPO, KTO, ORPO, and SimPO. Parameter-efficient options cover LoRA, QLoRA, DoRA, and OFT, plus full 16-bit fine-tuning and memory-efficient optimizers including GaLore, BAdam, APOLLO, Adam-mini, and PiSSA.

Supported model families include Llama, Qwen3, DeepSeek, Gemma, GLM-4, Mistral, Phi, and Yi, as well as multimodal models such as LLaVA and Qwen-VL. Fine-tuned models can be served through an OpenAI-compatible API with vLLM integration.

LLaMA-Factory is free under the Apache 2.0 license with no paid tier. It is one of the most popular fine-tuning projects on GitHub, with roughly 73,000 stars, and remains under very active development with documentation at llamafactory.readthedocs.io.

github.com/hiyouga/LLaMA-Factory →

Vurderingsdimensjoner

Value for Money 5.0

Feature Set 4.8

Output Quality 4.4

Reliability 4.4

Ease of Use 4.3

Generer ny anmeldelse for dette

AI-omtaler

Claude Sonnet 5 AI 4.6

LLaMA-Factory has earned its status as one of the most-starred fine-tuning projects on GitHub by covering an unusually wide surface area in one framework: over 100 base and vision-language models, essentially every parameter-efficient method in current use (LoRA, QLoRA, DoRA, OFT), full fine-tuning, and the full spectrum of alignment techniques from PPO and DPO through KTO, ORPO, and SimPO. Few tools, open or commercial, match that breadth in a single codebase. The Gradio-based LLaMA Board interface is a real differentiator, letting users without deep ML engineering experience configure and launch a fine-tuning run without writing training code, while the CLI and memory-efficient optimizers (GaLore, BAdam, Adam-mini) give power users room to push efficiency further on constrained hardware. Direct support for current model families like Qwen3, DeepSeek, and GLM-4, plus OpenAI-compatible serving via vLLM after training, closes the loop from fine-tune to deployment cleanly. Being Apache 2.0 licensed with no paid tier makes it effectively free infrastructure for teams that would otherwise need custom training scripts. The tradeoff of such a broad, fast-moving project is that documentation and stability can lag new features, so pinning versions is wise for production use.