Core Libraries
| Library | Purpose | Key API |
|---|
| transformers | Model loading + inference + training | pipeline('text-generation'), AutoModel, AutoTokenizer |
| datasets | Dataset loading + processing | load_dataset(), map(), filter(), train_test_split() |
| accelerate | Multi-GPU/distributed training | Accelerator(), notebook_launcher() |
| peft | Parameter-efficient fine-tuning | LoraConfig, get_peft_model(), PeftModel |
| trl | RLHF, DPO, reward modeling | SFTTrainer, DPOTrainer, RewardTrainer |
| diffusers | Image/audio generation | StableDiffusionPipeline, DiffusionPipeline |
| safetensors | Safe model format (no pickle) | save_file(), load_file() |
Model Discovery
| Item | Description |
|---|
Model Card | README equivalent — architecture, training data, benchmarks, limitations. Read before using. |
Downloads & Likes | Proxy for quality. High downloads + likes = community-validated. |
Tags & Filters | Filter by task (text-generation), library (transformers), language, license. |
Leaderboards | Open LLM Leaderboard, LMSys Chatbot Arena — compare models on standardized benchmarks. |
GGUF Models | Quantized for local use — search 'GGUF' tag. TheBloke and QuantFactory are top uploaders. |
Model Variants | Check 'files' tab: safetensors vs pytorch, fp16 vs fp32, GGUF quantizations. |
Spaces & Deployment
| Item | Description |
|---|
Gradio Spaces | Interactive demos — drag-and-drop UI. Free CPU, paid GPU. Great for showcasing models. |
Static Spaces | Host HTML/JS/CSS — documentation, model demos without backend. |
Docker Spaces | Full container — custom dependencies, APIs. Most flexible. |
Inference Endpoints | Managed deployment — auto-scaling, per-hour pricing. Production-ready. |
Inference API (Serverless) | Pay-per-token. Good for prototyping, not production scale. |
ZeroGPU | Free GPU access for Spaces — shared, queued. Great for open-source demos. |
Best Practices
| Item | Description |
|---|
Use safetensors | Never pickle — security risk. Most new models ship in safetensors format. |
Pin versions | transformers==4.40.0 — breaking changes happen. Pin in requirements.txt. |
Check license | Apache/MIT = free use. RAIL = responsible use. CC-BY-NC = non-commercial only. |
trust_remote_code=False | Don't run arbitrary code from untrusted models. Review modeling code first. |
Model caching | Models downloaded to ~/.cache/huggingface/ — reuse across projects. Set HF_HOME to change. |
Pro Tip: The HF Hub is GitHub for ML — 500K+ models, 100K+ datasets. Use the 'like' and filter system to find quality models: sort by downloads, likes, and check the model card for evaluation results.