Hugging Face
The open model hub and Transformers ecosystem — I version my ML datasets there, and studied its libraries in depth in the IBM certificate.
Hugging Face hosts open models and datasets and maintains libraries such as Transformers, Datasets and TRL, used for loading, fine-tuning and serving models.
How I use it
- Real-time Data Systems — I version the datasets for my ML research on Hugging Face; Google Colab notebooks pull them to fine-tune IBM Granite TSPulse, a time-series foundation model, with PyTorch.
- local-rag-chatbot — small sentence-transformers embedding models that run on CPU.
- Coursework — in the IBM Generative AI Engineering Professional Certificate professional certificate, Hugging Face libraries are used for data loading, Transformers, instruction tuning and preference optimization (DPO) — see Generative AI and LLMs: Architecture and Data Preparation, Generative AI Advanced Fine-Tuning for LLMs and AI Agents Using RAG and LangChain. LLM fine-tuning with these libraries remains coursework for me rather than shipped work.
Why it matters
Understanding the open-model stack — tokenizers, model hubs, training loops — informs the trade-offs I make at the application layer, such as when a small local embedding model is good enough and when a hosted model is worth its cost.