The Hugging Face Stack

~18 min · transformers, trl, peft, datasets, accelerate

Level 0Observer

0 XP0/43 lessons0/11 achievements

0/120 XP to next level120 XP to go0% complete

Five libraries, one pipeline

The open-source fine-tuning ecosystem centers on five Hugging Face libraries that compose into one clean pipeline.

Library	Purpose	2025–2026 version
transformers	Model loading, tokenization, inference.	4.x
datasets	Data loading, processing, formatting.	3.x
peft	LoRA, QLoRA, DoRA, and other PEFT methods.	0.17+
trl	Training loops (SFTTrainer, DPOTrainer, GRPOTrainer).	1.0
accelerate	Multi-GPU, distributed training.	1.x

TRL v1.0 (March 2026) consolidated SFTTrainer, DPOTrainer, GRPOTrainer, and more into a stable API. Key change: TrainingArguments and SFTConfig were unified — SFTConfig now extends TrainingArguments with all SFT-specific parameters in one place.

Code

Install the full stack·bash

# Install everything
pip install torch transformers datasets peft trl accelerate bitsandbytes

# Optional: Weights & Biases experiment tracking
pip install wandb

# Verify versions
python -c "import transformers, peft, trl; print(transformers.__version__, peft.__version__, trl.__version__)"

External links

Exercise

Set up a fresh Python environment (venv or conda) and install all five libraries. Verify each imports without error and print its version. Pin those exact versions in a requirements.txt. This becomes your reproducible base for every fine-tuning project.

Progress

Progress is local-only — sign in to sync across devices.

← 🧩 PEFT and AdaptersQuiz · 5 questions Next →Loading Models & Tokenizers

Spotted a bug or have feedback on this page?Report an Issue

Comments 0

🔔 Reply notifications (sign in)

No comments yet — be the first.