🤗 HuggingFace Project Learning

Hands-on work from my journey through the Hugging Face LLM Course and The Reasoning Course, plus a flagship end-to-end project: fine-tuning BERT for Romanian extractive question answering on the XQuAD-ro dataset.

Companion to my GitHub profile — the HF completion certificates (LLM Course Unit 1, Unit 3, and The Reasoning Course — Fundamentals of GRPO) shown there were earned working through the notebooks in this repo.

---

🚀 Flagship Project — Romanian QA on XQuAD-ro

Two extractive QA models fine-tuned on the Romanian split of XQuAD, loaded directly from the DeepMind mirror as xquad.ro.json and prepared with 🤗 datasets.

Notebook	Base model	Approach	Colab
bert-base-romanian-cased-v1.ipynb	dumitrescustefan/bert-base-romanian-cased-v1	Monolingual Romanian BERT
multilingual-bert.ipynb	bert-base-multilingual-cased	Multilingual BERT baseline

Both notebooks cover the full pipeline: dataset download → tokenization with sliding-window context → AutoModelForQuestionAnswering fine-tuning → post- processing spans → EM / F1 evaluation. See projects/romanian-qa-xquad-ro/README.md for the model comparison.

---

🧠 Reasoning & GRPO (→ Reasoning Course certificate)

Group Relative Policy Optimization — the RL algorithm behind reasoning models — studied three ways: from scratch in raw PyTorch, with 🤗 trl, and with unsloth for memory-efficient training.

Notebook	What it covers
GRPO_Implementation_in_PyTorch.ipynb	GRPO from first principles on Qwen/Qwen2-Math-1.5B — group sampling, reward computation, advantage normalization and the policy update, written out by hand
Fine_tuning_a_model_with_GRPO.ipynb	GRPOTrainer from trl + LoRA (peft) on HuggingFaceTB/SmolLM-135M-Instruct over the mlabonne/smoltldr summarization set, tracked in Weights & Biases
nb/Gemma3_(1B)-GRPO.ipynb	google/gemma-3-1b-it on GSM8K with unsloth + vllm — custom reward functions for answer format and correctness, then pushing the result to the Hub

---

📚 HF LLM Course Notebooks

Selected chapters from the Hugging Face LLM course, reworked with my own experiments and notes.

Chapter 3 — Fine-tuning a pretrained model (→ Unit 3 certificate)

Task: GLUE / MRPC paraphrase classification with bert-base-uncased.

• 01_trainer_api.ipynb — Trainer API end-to-end
• 02_full_training_loop.ipynb — same task with a hand-written PyTorch loop + accelerate
• 03_learning_curves.ipynb — reading and debugging training curves
• Evaluation.ipynb — metrics with evaluate

Chapter 4 — Sharing models and tokenizers

• 01_using_pretrained_models.ipynb
• 02_using_and_sharing.ipynb — pushing checkpoints to the Hub

Chapter 5 — The 🤗 Datasets library

• 01_importing_external_datasets.ipynb — CSV / JSON / local files
• 02_slicing_and_dicing.ipynb — map, filter, train_test_split on UCI drugsCom reviews
• 03_handling_big_datasets.ipynb — streaming + memory-mapping PubMed summarization
• Semantic_Search_with_FAISS.ipynb — embedding the lewtun/github-issues corpus and building a FAISS index

Chapter 6 — The 🤗 Tokenizers library

• Fast_Tokenizers_Special_Powers.ipynb — offset mappings, word IDs, QA/NER pipelines
• Training_a_New_Tokenizer_from_an_Old_One.ipynb — retraining on code_search_net
• Normalization_and_Pre-tokenization.ipynb — the pieces of the tokenization pipeline
• Building_A_Tokenizer_from_Scratch.ipynb — assembling one block by block
• WordPiece_from_Scratch.ipynb — the WordPiece algorithm by hand
• Unigram_from_Scratch.ipynb — the Unigram algorithm by hand

Chapter 7 — Classic NLP tasks

Each notebook fine-tunes and pushes a checkpoint to the Hub.

• TokenClassification.ipynb — NER on CoNLL-2003 → bert-finetuned-ner
• Fine-tuning_a_Masked_Language_Model.ipynb — domain-adapting DistilBERT to IMDB
• Translation.ipynb — Helsinki-NLP/opus-mt-en-fr on KDE4 en→fr
• Summarization.ipynb — google/mt5-small on multilingual Amazon reviews
• Question_answering.ipynb — BERT on SQuAD (the English precursor to the XQuAD-ro project above)
• Training_a_Data_Science_Syntax_Auto_Completer_Model(Causal Language Model)_from_Scratch.ipynb — a GPT-2-style code completer trained from scratch on CodeParrot

Chapter 8 — How to ask for help

• error_fix_workflow.ipynb — debugging the training pipeline and reading tracebacks

Chapter 9 — Building demos with Gradio

• Gradio_Interference_Class.ipynb — the Interface basics
• Advanced_Interface_features.ipynb — state, interpretation, layout
• Playing_with_the_Gradio_API.ipynb — driving a Space programmatically
• Gradio_Integration_with_HuggingFace.ipynb — wiring demos to Hub models and Spaces

Chapter 11 — Fine-tuning LLMs

• Chat_Templates.ipynb — chat formats across SmolLM2 / Qwen / Mistral, converting smoltalk to model-ready text
• Supervised Fine-Tuning with SFTTrainer.ipynb — SFTTrainer on HuggingFaceTB/SmolLM2-135M
• LoRA SFT.ipynb — parameter-efficient SFT with peft adapters, then merging them back
• Evaluation.ipynb — benchmarking with lighteval

---

🛠️ Stack

Core transformers · datasets · evaluate · accelerate · torch · tokenizers

Fine-tuning & RL trl · peft · unsloth · bitsandbytes · vllm

Tooling gradio · lighteval · faiss · wandb · scikit-learn

▶️ Running locally

pip install -r requirements.txt
jupyter lab

requirements.txt covers the core course notebooks. The GRPO, SFT and evaluation notebooks pull their heavier extras (trl, peft, unsloth, vllm, lighteval, wandb) inline with pip install in the first cell, so they run as-is in Colab.

Every notebook also carries an Open in Colab badge — a free GPU runtime is enough for most course notebooks; the QA fine-tuning and GRPO notebooks want a T4 or better.

🗂️ Repo layout

projects/   Flagship end-to-end projects (Romanian QA)
course/     HF course chapter walkthroughs (ch3–ch9, ch11)
nb/         Standalone notebooks (Gemma 3 GRPO)
*.ipynb     Reasoning course GRPO notebooks

🔗 More

• Author: @pop123-ux
• Medium write-ups: medium.com/@Pop123

📄 License

MIT.