anurag629's journal - RepoJournal

Anurag converted the gen_ai_engineer course to Jupyter notebooks with Colab support, cleaned up the duplicates and stale references, then reverted the entire change back to markdown and Python files before adding a transformer implementation for Day 3.

The work began in anurag629/gen_ai_engineer with a redesign of the README for GitHub Pages [1], adding shield badges for Python, PyTorch, Hugging Face and LangChain, a navigation bar with anchor links, and a progress tracker table with direct links to all 21 day directories. Day sections were converted to HTML tables for card-style layout, and raw URLs became proper markdown hyperlinks.

Then came a major format shift: converting all content to Jupyter notebooks with Colab badges [2]. Day 1-3 README books and all Python files (micrograd, exercises, visualizations, bigram, bigram_neural, mlp_lm) were converted to .ipynb format. Each notebook included an "Open in Colab" badge, auto-installed dependencies, and downloaded required data files for Colab. Image paths were replaced with GitHub raw URLs for display compatibility.

Following this, Anurag cleaned house by removing duplicated files and keeping only .ipynb as the primary format [3]. Three day README files, seven .py files, and three runtime-generated PNGs were deleted. The exercises.ipynb was made self-contained by inlining micrograd classes, and a convert_to_notebooks.py script was added for regeneration.

Stale references were then fixed [4]: old `python micrograd.py` and `python visualizations.py` commands were removed from Day 01 notebooks, broken indentation in visualizations.ipynb was corrected, and .gitkeep files were added to 18 empty day directories so they would appear on GitHub.

Then the entire notebook conversion was reverted [5]. All .ipynb files were removed and the original README.md books and .py source files were restored. The convert_to_notebooks.py script was deleted, and the main README was updated to link back to .md and .py files instead of notebooks, with Colab badges removed.

Finally, Anurag added a Day 3 mini GPT implementation [6]. The mini_gpt.py file contains a complete decoder-only Transformer language model trained on the names dataset, using GPT-style training with LR warmup, cosine decay, gradient clipping, and causal masking. An MLP baseline comparison shows the Transformer achieves lower validation loss (2.02 vs 2.08). The implementation symlinks names.txt from Day 2.

Anurag Verma

Notebook conversion and revert cycle, mini GPT on Day 3

Notebook conversion and revert cycle, mini GPT on Day 3

Follow anurag629