The user pop123-ux has published a learning repository containing 38 runnable notebooks designed to help users understand the Hugging Face stack by progressively removing high-level abstractions. The collection covers a path from core components like transformers and tokenizers into fine-tuning, PEFT, reasoning/RL, and end-to-end project work.
- Includes implementations of WordPiece and Unigram tokenizers built from scratch.
- Demonstrates LoRA and supervised fine-tuning alongside GRPO implemented from first principles before using GRPOTrainer.
- Covers FAISS semantic search, trl, Unsloth, and vLLM experiments.
- Features a Romanian XQuAD project comparing monolingual and multilingual BERT with corrected article-level held-out evaluation.
The author seeks technical feedback on the clarity of the GRPO-from-scratch and Romanian QA evaluation sections, as well as suggestions for gaps in the learning path.