Colab AI Project Collection

These notebooks were developed in 2024 for learning and experimentation—exploring NLP and modern LLMs end-to-end rather than serving production systems.

A portfolio of hands-on NLP and transformer notebooks, originally built in Google Colab, covering classical text pipelines through modern LLM fine-tuning with Hugging Face Transformers, parameter-efficient adaptation (LoRA), and standard evaluation metrics.

Each notebook is self-contained and includes an Open in Colab badge pointing at this repository.

---

What’s inside (project-by-project)

Notebook	What you explore	Stack & datasets (high level)
Text_Cleaning_Project.ipynb	End-to-end text normalization for downstream NLP: regex cleanup, punctuation, chat-word handling, light spell correction, stop-word removal, and emoji stripping, on real review text.	pandas; IMDB sentiment corpus (CSV via URL in notebook)
Word2Vec_Project.ipynb	Train a Word2Vec model from scratch with Gensim, preprocess long-form fiction text, and inspect similarity and doesn’t-match behavior—classic static embeddings before transformers.	NLTK, Gensim; Game of Thrones books (Kaggle-style layout referenced in notebook)
Sentence_Embedding_Project.ipynb	BERT-based sentence embeddings: tokenization with bert-base-uncased, forward pass through BertModel, pooling/context vectors, and decoding back to readable tokens.	Hugging Face Transformers; PyTorch/CUDA where available
NER_with_BERT_Project.ipynb	Named Entity Recognition as token classification: fine-tune bert-base-uncased with AutoModelForTokenClassification, aligned token–label encoding, training with Trainer, and evaluation (seqeval / evaluate). Optional push to Hugging Face Hub.	datasets (CoNLL-2003), Transformers, Accelerate
Q&A_with_LM.ipynb	Extractive question answering on SQuAD-style data: predict an answer span in the passage (or “unanswerable” behavior as supported by the pipeline). Uses a BERT-style reader via SimpleTransformers for fast training/eval.	SimpleTransformers; bert-base-cased; Stanford QA data (paths in notebook)
Text_Summarization_Project.ipynb	Abstractive summarization with a seq2seq transformer: start from Pegasus (google/pegasus-cnn_dailymail), fine-tune on dialogue → summary data (Samsung/samsum), use TrainingArguments + Trainer, and measure quality with ROUGE via evaluate.	Transformers, Datasets, ROUGE
Llama2_LoRA_Project.ipynb	Instruction tuning for Llama 2 7B with LoRA (PEFT), quantization-friendly tooling (bitsandbytes where applicable), and TRL-style supervised fine-tuning on an instruction dataset (mlabonne/guanaco-llama2-1k). Illustrates efficient LLM adaptation vs full fine-tuning.	PEFT, TRL, Accelerate, Transformers

---

Skills and themes reflected in this repo

These notebooks map cleanly to common hiring and portfolio keywords:

• Natural Language Processing: cleaning, tokenization, embeddings, span QA, NER, summarization
• Transformers & LLMs: BERT (encoder), Pegasus (encoder–decoder), Llama 2 (decoder)
• Fine-tuning: full fine-tuning (NER, summarization), LoRA / PEFT (Llama 2)
• Hugging Face ecosystem: transformers, datasets, evaluate, Hub workflows
• Model evaluation: span metrics for QA/NER, ROUGE for summarization
• Efficient training: mixed precision / Trainer patterns, parameter-efficient LLM tuning

GPT-class APIs, LangChain, RAG, AWS Bedrock, and Azure AI are important parts of the modern LLM stack but are not implemented as standalone projects in this repository. If you use those in production or other repos, call them out on your profile or add a short “Elsewhere” section there—keep this README tightly aligned with what visitors can actually run from these notebooks.

---

Tech stack (from the notebooks)

Python · Jupyter / Google Colab · PyTorch · Hugging Face (transformers, datasets, accelerate, evaluate, peft, trl) · NLTK · Gensim · SimpleTransformers (QA) · ROUGE · seqeval (NER)

---

How to run

1. Recommended: open any notebook on Google Colab via the badge at the top of the file (points to jackychh7878/Colab_AI_Project on GitHub).
2. Locally: use Python 3.10+ (as in the Colab logs), install dependencies per notebook (pip install … cells), and ensure GPU availability for larger models (BERT fine-tunes, Pegasus, Llama 2 + LoRA).
3. Secrets: notebooks that use the Hugging Face Hub may expect a token (notebook_login / HF_TOKEN patterns)—add your own token where prompted.
4. GitHub shows “Invalid Notebook” (missing application/vnd.jupyter.widget-state+json): Colab/Hugging Face often save tqdm output as ipywidgets. GitHub’s viewer is strict about widget JSON. From the repo root run python fix_notebooks_for_github.py — it removes only the widget MIME bundle from each output and keeps text/plain (and images, etc.), so you do not need to re-run the notebook.

---

Repository layout

Colab_AI_Project/
├── Text_Cleaning_Project.ipynb
├── Word2Vec_Project.ipynb
├── Sentence_Embedding_Project.ipynb
├── NER_with_BERT_Project.ipynb
├── Q&A_with_LM.ipynb
├── Text_Summarization_Project.ipynb
├── Llama2_LoRA_Project.ipynb
└── README.md

---

Reference

• Author: Jacky Chong — Colab AI project collection.
• Medical chatbot (Retrieval-Augmented Generation) — write-up / notes: Google Doc
• Google Drive backup of materials: folder

---