🤗 Hugging Face Learning Journey

This repository documents my step-by-step learning of Hugging Face tools for applied machine learning, with a focus on text data and real-world workflows.

The goal is to build a clear understanding of the full pipeline:

data → preprocessing → training → evaluation → sharing → application

01 - Loading Data

What this covers:

loading datasets using load_dataset accessing train/test splits initial dataset exploration

02 - Inspect Dataset

What this covers:

inspecting samples and features understanding dataset structure checking dataset size and label distribution

03 - Preprocessing

This notebook covers:

• general preprocessing such as filtering and text transformation
• text preprocessing for transformers
• tokenization with a matching pretrained tokenizer
• preparing the dataset for PyTorch

04 - Fine-Tuning With Trainer API

This notebook continues after preprocessing and shows how to:

• load a pretrained text classification model
• define training settings with TrainingArguments
• use Trainer to fine-tune the model
• evaluate performance with accuracy

05 - Custom Training Loop

This notebook explains how text classification training works without the Trainer API:

• prepare tokenized data for PyTorch
• create DataLoaders
• define optimizer and scheduler
• run a manual training loop
• evaluate the model

06 - Understanding Learning Curves

This notebook explains how to interpret:

• training and validation loss
• validation accuracy
• healthy learning curves
• overfitting
• underfitting
• unstable training

07 - Sharing Models

Use and share models via the Hugging Face Hub.

• loading pretrained models
• saving models locally
• pushing models to the Hub
• model cards

08 - Classical NLP Tasks

This notebook covers the most practical classical NLP tasks that form the foundation of modern language models:

• token classification
• question answering
• summarization
• translation

For each task, I included:

• the main idea
• a small practical example
• the key preprocessing concept
• the common evaluation metric
• a short summary of what I learned

Main takeaways

• token classification requires label alignment with subword tokens
• extractive QA requires mapping answer spans from characters to tokens
• summarization typically uses encoder-decoder models and ROUGE
• translation uses sequence-to-sequence models and SacreBLEU

9 - Supervised Fine-Tuning and LoRA

This notebook covers the practical workflow for adapting a pretrained language model into an assistant-style model.

Topics covered

• chat templates
• supervised fine-tuning (SFT)
• LoRA
• evaluation after fine-tuning

Main takeaways

• instruct models require the correct chat template
• SFT is useful only when prompting is not enough
• LoRA makes fine-tuning much more memory-efficient
• evaluation should include both metrics and qualitative checks