Kenhuang Academy
  • English

      • Hugging Face Transformers Tutorial

      Welcome to the Hugging Face tutorial. The homepage card is titled Hugging Face; the content focuses on the Transformers library and the Hub (models, datasets, Spaces). It follows the official Transformers docs and huggingface/transformers—from first inference to fine-tuning and sharing.

      Transformers is the model-definition framework for text, vision, audio, video, and multimodal models. The Hub hosts 1M+ checkpoints. Runnable examples use ungated small models so you are not blocked by Llama-style license walls.

      Table of Contents

      Basics

      1. Introduction — Hub, Transformers, the ecosystem, and how this course relates to PyTorch / Ollama
      2. Installation — venv, PyTorch, transformers, hf auth login; GPU optional
      3. Quick Start — Classification or generation with pipeline

      Core APIs

      1. Pipeline — Task table; text, vision, and audio examples
      2. Tokenizers and Models — AutoTokenizer, AutoModelForCausalLM, encode / decode
      3. Text Generation — generate(), greedy vs sampling, max_new_tokens

      Data and Training

      1. Datasets — load_dataset, map, DataCollator
      2. Fine-tuning — Trainer + rotten_tomatoes, or LoRA on a small causal LM
      3. The Hub — Search, model cards, push_to_hub, Spaces teaser

      Practice

      1. Practical Examples — Sentiment, a RAG-less chatbot, fine-tune outline
      2. Resources — Official docs and related courses on this site

      Learning path

      StageGoalChapters
      Day 1Run inferenceIntro → Install → Quick Start → Pipeline
      Days 2–3Weights and generationTokenizers and Models → Generation
      Week 1Data and fine-tuningDatasets → Fine-tune → Hub
      Week 2A small projectExamples → Resources

      Models used here (all publicly downloadable)

      UseModel / dataWhy
      Text classificationdistilbert/distilbert-base-uncased + rotten_tomatoesSmall, ungated, ideal for Trainer
      Off-the-shelf sentimentdistilbert/distilbert-base-uncased-finetuned-sst-2-englishCommon official pipeline example
      Generationdistilbert/distilgpt2 or Qwen/Qwen2.5-0.5B-InstructQwen is stronger in Chinese; Qwen / DeepSeek are CN-friendly
      Speech recognitionopenai/whisper-tinyASR you can try on CPU
      • Tensors and training loops: start with the PyTorch tutorial, then return for Trainer / generate().
      • Local chat only: the Ollama tutorial is easier; this course is for research, fine-tuning, and custom pipelines.
      • Coding agents: Hugging Face Skills (SKILL.md) can drive Hub operations—not the main path here.

      Prerequisites

      • Python 3.10+ (functions, virtual environments)
      • The idea of pretrain → fine-tune → infer; you do not need to implement Transformer blocks
      • A GPU helps; classification and tiny generation work on CPU

      Study tips

      1. Follow the chapter order; every page has runnable code
      2. First downloads fill the Hub cache—leave disk space
      3. If you hit a gated repo, switch to the ungated IDs listed above
      4. Read the model card (license, VRAM) before fine-tuning
      5. Coding agents can use Hugging Face Skills—run a Pipeline yourself first

      Next steps

      评论