// Software stack
Fine-tune open-weight models
Prepare data, fine-tune on modest GPUs, track experiments, then serve the result as an API.
6 pieces5 setup stepsscenario: Run models on your own machine
What goes in it
- 1 ★ 78k
unsloth Training & fine-tuning Open source
Fast, memory-efficient fine-tuning
Local UI to run and train LLMs and diffusion models. Supports GGUF, MLX, Qwen3.8, DeepSeek-V4, MiniMax-H3, Gemma 4, FLUX and more.
- 2 ★ 76k
LlamaFactory optional Training & fine-tuning Open source
All-in-one fine-tuning with a UI
Unified Efficient Fine-Tuning of 100+ LLMs & VLMs (ACL 2024)
- 3 ★ 22k
datasets Training & fine-tuning Open source
Loads and processes training data
🤗 The largest hub of ready-to-use datasets for AI models with fast, easy-to-use and efficient data manipulation tools
- 4 ★ 20k
trl Training & fine-tuning Open source
Post-training such as SFT and DPO
Train transformer language models with reinforcement learning.
- 5 ★ 11k
wandb Observability & tracing Open source
Experiment tracking
The AI developer platform. Use Weights & Biases to train and fine-tune models, and manage models from experimentation to production.
- 6 ★ 95k
vllm Local inference & serving Open source
Serves the model with high throughput
A high-throughput and memory-efficient inference and serving engine for LLMs
How to set it up
- 01Make sure you really need it: fine-tune only when prompting and RAG fall short
- 02Prepare data with datasets and hold out an evaluation set
- 03Use Unsloth for LoRA/QLoRA on one GPU, LLaMA-Factory for a UI, TRL for DPO and other post-training
- 04Track every run in W&B
- 05Merge weights, serve with vLLM and compare against the base model on the eval set
✓ Good fit
- Fixed formats, domain vocabulary, strong style requirements
✗ Look elsewhere if
- Knowledge changes often: RAG fits better