Hugging Face Quest

Tracks

1. 01🏠Hub Literacy

   0/6 lessons

   Repos, cards, tokens, caches

   The Hub is a Git-backed registry. Every model, dataset, and Space is a repo with a contract. Treat it that way and the rest of the ecosystem composes.

   Lesson list (6)

   • 01
     What Hugging Face Actually Is~30 min · hub, ecosystem
   • 02
     Repository Types: Models, Datasets, Spaces~28 min · hub, repos
   • 03
     Browsing, Filtering, Naming Conventions~26 min · hub, search
   • 04
     Model Cards & YAML Front-Matter~32 min · hub, model-cards
   • 05
     Tokens, Access, and CLI Login~24 min · hub, auth
   • 06
     Downloads, Caching, and Pinning~28 min · hub, cache

   ✓Quiz · 5 questions→

2. 02🤖Transformers Library

   0/6 lessons

   From pipeline() to generate()

   The Transformers library is one Auto* registry, three backends (PyTorch default), and a generation state machine. Build past pipeline() into the controllable layer underneath.

   Lesson list (6)

   • 01
     Installation and the Three-Backend Reality~25 min · transformers, install
   • 02
     The pipeline() Abstraction~28 min · transformers, pipeline
   • 03
     Tasks, Auto Classes, and from_pretrained()~30 min · transformers, auto
   • 04
     Tokenizers In Depth: BPE, WordPiece, SentencePiece~32 min · transformers, tokenization
   • 05
     Generation Configuration: Sampling, Beam, Constraints~30 min · transformers, generation
   • 06
     Saving, Loading, and push_to_hub~25 min · transformers, hub, save

   ✓Quiz · 5 questions→

3. 03📦Datasets Library

   0/8 lessons

   Load, transform, stream, push

   Datasets is Arrow under the hood, with a single load_dataset() that handles Hub repos, local files, and streaming sources. Map, filter, split, push to Hub — the whole loop.

   Lesson list (8)

   • 01
     load_dataset(): The Universal Loader~26 min · datasets, load
   • 02
     Dataset vs DatasetDict, Features, and Schema~24 min · datasets, schema
   • 03
     Exploring with Pandas, Arrow, and the Viewer~22 min · datasets, explore
   • 04
     dataset.map(): Transformations that Cache~30 min · datasets, map
   • 05
     Streaming Datasets and IterableDataset~26 min · datasets, streaming
   • 06
     Custom Datasets from Local Files~24 min · datasets, custom
   • 07
     Splitting, Combining, Concatenating~20 min · datasets, split
   • 08
     Caching, push_to_hub, and Reproducibility~24 min · datasets, hub, cache

   ✓Quiz · 5 questions→

4. 04🔌InferenceClient & Providers

   0/5 lessons

   One client, many backends

   InferenceClient is the abstraction that talks to HF-hosted, third-party providers, your own TGI box, and OpenAI-compatible endpoints. Same code, different backend with one string change.

   Lesson list (5)

   • 01
     InferenceClient: One Object, Many Backends~26 min · inference, client
   • 02
     Provider Routing and the Pricing Surface~24 min · inference, providers
   • 03
     Streaming, Tool Use, and Structured Output~30 min · inference, streaming
   • 04
     OpenAI-Compatible Endpoints~22 min · inference, openai-compat
   • 05
     Cost, Rate Limits, and Retry Strategy~26 min · inference, ops

   ✓Quiz · 5 questions→

5. 05🚀TGI and vLLM

   0/5 lessons

   Real serving, real concurrency

   Continuous batching, PagedAttention, serve-time quantization, health endpoints, and restart strategy. Move from pipeline()-behind-Flask to a server that handles real traffic.

   Lesson list (5)

   • 01
     Why a Dedicated Inference Server~24 min · serving, tgi, vllm
   • 02
     TGI: Pull, Run, Tune~30 min · serving, tgi
   • 03
     vLLM: PagedAttention and Continuous Batching~28 min · serving, vllm
   • 04
     Quantization at Serve Time~24 min · serving, quantization
   • 05
     Health Checks, Metrics, and Restart Strategy~22 min · serving, ops, monitoring

   ✓Quiz · 5 questions→

6. 06🎯Trainer + PEFT

   0/8 lessons

   SFT, LoRA, QLoRA, DPO

   Trainer + TrainingArguments + the right collator + the right metric is the canonical loop. PEFT (LoRA / QLoRA) is how you fine-tune big models on small GPUs. SFTTrainer wraps it for chat data.

   Lesson list (8)

   • 01
     Trainer and TrainingArguments~30 min · training, trainer
   • 02
     Data Collators: Padding and Beyond~22 min · training, collator
   • 03
     Evaluation Metrics and the evaluate Library~24 min · training, eval, metrics
   • 04
     Mixed Precision, Gradient Accumulation, and Memory~26 min · training, memory, amp
   • 05
     Why Fine-Tune (vs. Prompt, vs. RAG)~20 min · training, decision
   • 06
     LoRA: Low-Rank Adaptation~28 min · training, peft, lora
   • 07
     QLoRA and the PEFT Library~26 min · training, peft, qlora
   • 08
     SFTTrainer, Chat Templates, and Merging~26 min · training, trl, sft

   ✓Quiz · 5 questions→

7. 07🖼️Spaces and Apps

   0/4 lessons

   Gradio, Streamlit, Docker

   A Space is a Git repo with a manifest that runs an app. Gradio for ML demos, Streamlit for stateful dashboards, Docker for everything else, static for plain HTML.

   Lesson list (4)

   • 01
     What Spaces Actually Are~22 min · spaces, deploy
   • 02
     Gradio: ML Demos in 10 Lines~26 min · spaces, gradio
   • 03
     Streamlit, Docker, and Static Spaces~24 min · spaces, streamlit, docker
   • 04
     Secrets, Hardware Tiers, and Persistent Storage~24 min · spaces, ops, secrets

   ✓Quiz · 5 questions→

8. 08🛠️Ecosystem Operations

   0/8 lessons

   Safetensors, GGUF, MLX, DPO, E2E

   Safetensors security, quantization formats (AWQ/GPTQ/GGUF/MLX/ONNX), Accelerate + FSDP, Optimum's hardware runtimes, TRL for preference alignment, and the end-to-end Hub→Train→Serve→Space loop.

   Lesson list (8)

   • 01
     Safetensors: Why It's the Default~22 min · ops, safetensors, security
   • 02
     Quantization Formats Cheat Sheet~24 min · ops, quantization
   • 03
     GGUF Export for Ollama / llama.cpp~24 min · ops, gguf, ollama
   • 04
     ONNX and MLX Conversion~22 min · ops, onnx, mlx
   • 05
     Accelerate: Multi-Device, Mixed Precision, FSDP~22 min · ops, accelerate
   • 06
     Optimum: Hardware-Specific Optimization~20 min · ops, optimum
   • 07
     TRL: RLHF, DPO, and Preference Optimization~24 min · ops, trl, rlhf
   • 08
     End-to-End: Hub → Train → Serve → Space~28 min · ops, pipeline

   ✓Quiz · 5 questions→