Hugging Face Ecosystem Cheat Sheet

Updated 2026-04-28

Next Topic: Hugging Face Transformers Cheat Sheet

🧠Study flashcards on this topic131 cards · spaced repetition→

Hugging Face is an open-source platform and community for machine learning that provides libraries, tools, model repositories, and infrastructure for building, training, sharing, and deploying AI models. The ecosystem encompasses the Transformers v5 library for state-of-the-art models (now PyTorch-first with modular architecture), the Hub for hosting 2M+ models and 500k+ datasets, Datasets for data processing, smolagents for AI agents, and dozens of specialized libraries covering NLP, computer vision, diffusion, reinforcement learning, and robotics. What distinguishes Hugging Face is its radically accessible design—complex ML workflows are abstracted into simple APIs while retaining full configurability, making cutting-edge AI practical for practitioners at every level, from prototyping with single-line pipelines to deploying production inference endpoints at scale.

What This Cheat Sheet Covers

This topic spans 16 focused tables and 138 indexed concepts, 131 flashcards. Below is a complete table-by-table outline of this topic, spanning foundational concepts through advanced details.

Table 1: Core LibrariesTable 2: Hugging Face HubTable 3: Transformers Library — Core ClassesTable 4: Model Loading and ConfigurationTable 5: Chat TemplatesTable 6: Dataset OperationsTable 7: Training ConfigurationTable 8: Fine-Tuning ApproachesTable 9: Quantization and CompressionTable 10: Inference and GenerationTable 11: Deployment ToolsTable 12: Model MergingTable 13: Advanced LibrariesTable 14: Hub UtilitiesTable 15: Community and CollaborationTable 16: Security and Safety

Quick IndexSubscribe to unlock

A jump-to index of every table row in this cheat sheet.

Mind MapSubscribe to unlock

An interactive map of every table and concept in this topic.

Table 1: Core Libraries

These are the workhorse packages most Hugging Face projects pull in, each owning one slice of the ML lifecycle — Transformers for models, Datasets for data, PEFT and TRL for fine-tuning and alignment, Diffusers for generation, Accelerate for scaling. Skim this first to know which library to reach for before diving into any task.

Library	Example	Description
Transformers	`from transformers import pipeline` `classifier = pipeline("sentiment-analysis")` `classifier("This is amazing!")`	• Flagship library providing 400+ pretrained transformer model architectures for NLP, vision, audio, and multimodal tasks • v5 is PyTorch-first (drops TensorFlow/Flax), introduces modular architecture, `transformers serve` OpenAI-compatible server, and quantization as a first-class citizen.
Datasets	`from datasets import load_dataset` `ds = load_dataset("glue", "sst2")` `ds["train"][0]`	• Library for efficient loading and processing of datasets with Apache Arrow backend • supports streaming, mapping, filtering, caching, and memory-mapped access for datasets larger than RAM.
PEFT	`from peft import LoraConfig, get_peft_model` `config = LoraConfig(r=8, lora_alpha=32)` `model = get_peft_model(model, config)`	• Parameter-Efficient Fine-Tuning library implementing LoRA, QLoRA, DoRA, AdaLoRA, IA3, Prefix Tuning, and P-Tuning • trains adapters that modify <1% of parameters while achieving comparable performance to full fine-tuning.
TRL	`from trl import SFTTrainer, GRPOTrainer` `trainer = SFTTrainer(model, args, train_dataset)` `trainer.train()`	• v1.0 post-training library with 75+ methods: SFT, DPO, GRPO, ORPO, KTO, PPO, RLOO, reward modeling • stable core (SFT, DPO, GRPO, RLOO) + experimental layer; integrates with vLLM for async generation.
Diffusers	`from diffusers import StableDiffusionPipeline` `pipe = StableDiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-xl-base-1.0")` `pipe("A cat in space")`	• State-of-the-art diffusion models for image, video, and audio generation • provides modular pipelines for Stable Diffusion XL, FLUX, ControlNet, InstructPix2Pix, and text-to-video models.
Accelerate	`from accelerate import Accelerator` `accelerator = Accelerator()` `model, optimizer = accelerator.prepare(model, optimizer)`	• Simplifies distributed training across any hardware configuration (multi-GPU, TPU, mixed precision) by adding just 4 lines of code to native PyTorch • handles device placement, gradient sync, and data parallelism automatically.
Tokenizers	`from transformers import AutoTokenizer` `tokenizer = AutoTokenizer.from_pretrained("gpt2")` `tokens = tokenizer("Hello world")`	• Rust-based fast tokenizers that train new vocabularies and tokenize at extreme speed (1GB in <20 seconds) • supports BPE, WordPiece, Unigram, and SentencePiece algorithms; v5 uses this as the sole tokenization backend.
smolagents	`from smolagents import CodeAgent, InferenceClientModel` `agent = CodeAgent(tools=[], model=InferenceClientModel())` `agent.run("Calculate sum 1 to 100")`	• Minimal library for AI agents in ~1000 lines of code; `CodeAgent` writes Python code to invoke tools and orchestrate other agents • model-agnostic (HF Hub, OpenAI, Anthropic), supports MCP tools, secure sandboxed execution.
Evaluate	`import evaluate` `accuracy = evaluate.load("accuracy")` `accuracy.compute(predictions=[0,1], references=[0,1])`	• Unified library providing 100+ evaluation metrics for NLP, vision, audio, and multimodal tasks • supports BLEU, ROUGE, BERTScore, perplexity, and task-specific metrics with a consistent API.
Optimum	`from optimum.onnxruntime import ORTModelForSequenceClassification` `model = ORTModelForSequenceClassification.from_pretrained("distilbert-base-uncased")`	• Hardware-specific optimization library for ONNX Runtime, Intel OpenVINO, Nvidia TensorRT, and other accelerators • exports models to optimized formats and applies quantization, pruning, and graph optimization.
Safetensors	`from safetensors.torch import save_file, load_file` `save_file(tensors, "model.safetensors")`	• Secure and fast tensor serialization format replacing pickle • prevents arbitrary code execution, supports lazy loading, and loads faster than .bin with zero-copy deserialization; default format on the Hub.

Table 2: Hugging Face Hub

The Hub is the GitHub of machine learning — the central place where models, datasets, and demos are hosted, versioned, and shared. The features here span discovery (Model and Dataset Hubs), the round-trip methods that load and publish your work (from_pretrained, push_to_hub), and the hosting surfaces (Spaces, Inference Endpoints) that turn a checkpoint into something others can run.

Feature	Example	Description
Model Hub	`Browse models at huggingface.co/models` `Filter by task, library, language`	• Central repository hosting 2M+ models including Llama, Qwen, DeepSeek, Mistral, Gemma, BERT, Stable Diffusion • supports version control via Git, model cards for documentation, and automated metrics display.

The rest of this section is part of the full cheat sheet

Table 3: Transformers Library — Core Classes

Almost everything you do in Transformers flows through this small set of classes. The Auto* family is the central trick — it inspects a checkpoint and instantiates the right model or tokenizer for you, so the same three lines work across hundreds of architectures, while pipeline wraps the whole thing for one-line inference and Trainer handles the training loop.

Class	Example	Description
pipeline	`pipe = pipeline("text-generation", model="gpt2")` `pipe("Once upon a time", max_length=50)`	• Highest-level API that wraps model, tokenizer, and post-processing into single call • supports 20+ tasks (classification, NER, QA, generation, translation) with automatic model selection.

The rest of this section is part of the full cheat sheet

Table 4: Model Loading and Configuration

from_pretrained accepts a surprising number of arguments, and the right ones decide whether a large model fits in your VRAM at all. These options cover the practical levers — picking a dtype like bf16, spreading layers across GPUs with device_map, loading quantized GGUF files, pinning a specific revision, and the trust_remote_code flag you should always think twice about.

Technique	Example	Description
Load from Hub	`model = AutoModel.from_pretrained("bert-base-uncased")`	• Downloads model from Hugging Face Hub to cache directory (`~/.cache/huggingface/`) • subsequent loads use cached version for offline access.

The rest of this section is part of the full cheat sheet

Table 1: Core Libraries

Library	Example	Description
Transformers	`from transformers import pipeline` `classifier = pipeline("sentiment-analysis")` `classifier("This is amazing!")`	• Flagship library providing 400+ pretrained transformer model architectures for NLP, vision, audio, and multimodal tasks • v5 is PyTorch-first (drops TensorFlow/Flax), introduces modular architecture, `transformers serve` OpenAI-compatible server, and quantization as a first-class citizen.
Datasets	`from datasets import load_dataset` `ds = load_dataset("glue", "sst2")` `ds["train"][0]`	• Library for efficient loading and processing of datasets with Apache Arrow backend • supports streaming, mapping, filtering, caching, and memory-mapped access for datasets larger than RAM.
PEFT	`from peft import LoraConfig, get_peft_model` `config = LoraConfig(r=8, lora_alpha=32)` `model = get_peft_model(model, config)`	• Parameter-Efficient Fine-Tuning library implementing LoRA, QLoRA, DoRA, AdaLoRA, IA3, Prefix Tuning, and P-Tuning • trains adapters that modify <1% of parameters while achieving comparable performance to full fine-tuning.
TRL	`from trl import SFTTrainer, GRPOTrainer` `trainer = SFTTrainer(model, args, train_dataset)` `trainer.train()`	• v1.0 post-training library with 75+ methods: SFT, DPO, GRPO, ORPO, KTO, PPO, RLOO, reward modeling • stable core (SFT, DPO, GRPO, RLOO) + experimental layer; integrates with vLLM for async generation.
Diffusers	`from diffusers import StableDiffusionPipeline` `pipe = StableDiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-xl-base-1.0")` `pipe("A cat in space")`	• State-of-the-art diffusion models for image, video, and audio generation • provides modular pipelines for Stable Diffusion XL, FLUX, ControlNet, InstructPix2Pix, and text-to-video models.
Accelerate	`from accelerate import Accelerator` `accelerator = Accelerator()` `model, optimizer = accelerator.prepare(model, optimizer)`	• Simplifies distributed training across any hardware configuration (multi-GPU, TPU, mixed precision) by adding just 4 lines of code to native PyTorch • handles device placement, gradient sync, and data parallelism automatically.
Tokenizers	`from transformers import AutoTokenizer` `tokenizer = AutoTokenizer.from_pretrained("gpt2")` `tokens = tokenizer("Hello world")`	• Rust-based fast tokenizers that train new vocabularies and tokenize at extreme speed (1GB in <20 seconds) • supports BPE, WordPiece, Unigram, and SentencePiece algorithms; v5 uses this as the sole tokenization backend.
smolagents	`from smolagents import CodeAgent, InferenceClientModel` `agent = CodeAgent(tools=[], model=InferenceClientModel())` `agent.run("Calculate sum 1 to 100")`	• Minimal library for AI agents in ~1000 lines of code; `CodeAgent` writes Python code to invoke tools and orchestrate other agents • model-agnostic (HF Hub, OpenAI, Anthropic), supports MCP tools, secure sandboxed execution.
Evaluate	`import evaluate` `accuracy = evaluate.load("accuracy")` `accuracy.compute(predictions=[0,1], references=[0,1])`	• Unified library providing 100+ evaluation metrics for NLP, vision, audio, and multimodal tasks • supports BLEU, ROUGE, BERTScore, perplexity, and task-specific metrics with a consistent API.
Optimum	`from optimum.onnxruntime import ORTModelForSequenceClassification` `model = ORTModelForSequenceClassification.from_pretrained("distilbert-base-uncased")`	• Hardware-specific optimization library for ONNX Runtime, Intel OpenVINO, Nvidia TensorRT, and other accelerators • exports models to optimized formats and applies quantization, pruning, and graph optimization.
Safetensors	`from safetensors.torch import save_file, load_file` `save_file(tensors, "model.safetensors")`	• Secure and fast tensor serialization format replacing pickle • prevents arbitrary code execution, supports lazy loading, and loads faster than .bin with zero-copy deserialization; default format on the Hub.

Table 2: Hugging Face Hub

Feature	Example	Description
Model Hub	`Browse models at huggingface.co/models` `Filter by task, library, language`	• Central repository hosting 2M+ models including Llama, Qwen, DeepSeek, Mistral, Gemma, BERT, Stable Diffusion • supports version control via Git, model cards for documentation, and automated metrics display.

Table 3: Transformers Library — Core Classes

Class	Example	Description
pipeline	`pipe = pipeline("text-generation", model="gpt2")` `pipe("Once upon a time", max_length=50)`	• Highest-level API that wraps model, tokenizer, and post-processing into single call • supports 20+ tasks (classification, NER, QA, generation, translation) with automatic model selection.

Table 4: Model Loading and Configuration

Technique	Example	Description
Load from Hub	`model = AutoModel.from_pretrained("bert-base-uncased")`	• Downloads model from Hugging Face Hub to cache directory (`~/.cache/huggingface/`) • subsequent loads use cached version for offline access.