Nick Swekosky (@NickSwekosky) header
Nick Swekosky (@NickSwekosky) profile picture

Senior PM @LumenTechCo. Exploring machine learning

Atlanta, GA
Joined February 2016
After reproducing LLM training from scratch and completing a fine-tuning exam, I’m pausing the Hugging Face course. Over Thanksgiving, I’ll sample ML engineering, deeper modeling, & LLM courses to choose a path toward useful projects. nickbuildsai.substack.com/p/…
11
This week, I worked through chapters 2 & 3 of the Hugging Face LLM course, exploring inference / fine-tuning while recognizing familiar training steps from fast.ai. Next, build working examples to understand what the pipeline & Trainer abstractions handle.
2
1
49
This week, I completed Chapter 1 of Hugging Face’s LLM course and diagrammed the differences between encoder-only, decoder-only, and encoder-decoder models. Next, I’ll start Chapter 2 and continue rebuilding fastai’s Road to the Top exercises from scratch nickbuildsai.substack.com/p/…
31
From @TKPPodcast Spite makes you work harder. Curiosity makes you work longer. Discipline is what you need when you have neither 🤔🤔🤔🤔
1
143
This week, I rebuilt my movie recommender from scratch and adapted it to predict restaurant ratings. Using only diner and restaurant IDs, my model had 29% lower validation mean squared error than predicting the training set’s average rating for everyone nickbuildsai.substack.com/p/…
18
This week, I rebuilt my housing-price model from scratch and explored how collaborative filtering predicts outcomes beyond movie ratings. nickbuildsai.substack.com/p/…
21
This week, I tested whether equal-weight ensembles of my random forest, linear model, & neural network could beat my best housing-price model; none did. Next, I’m rebuilding the workflow from memory before moving on to collaborative filtering. nickbuildsai.substack.com/p/…
17
This week, I trained a random forest with 25% lower error than my linear housing-price model on the same validation set. Then I used maps and feature importance to explore what it had learned about location and transit access. Link: nickbuildsai.substack.com/p/…
31
This week, I built and tested linear-model and neural-network ensembles to predict real-estate prices. Both had lower validation error than their members’ average, but neither beat its best individual model. nickbuildsai.substack.com/p/…
1
35
This week, I built a linear model and neural network to predict real-estate prices on an unfamiliar dataset. Along the way, I fixed mismatched error metrics and inconsistent feature scaling, and learned to rebuild/explain the workflow on my own nickbuildsai.substack.com/p/…
13
After six hours of waiting for my NLP classifier to train, it never finished; I cut the sample size way down, ran it again, and finished in a few minutes. In the future, I'll start with a smaller sample size and add more samples as needed. nickbuildsai.substack.com/p/…
14
This week, I trained a movie-review classifier to 94% training accuracy this week, using a sample of reviews from IMDB. Turns out, I was using a pretrained model without knowing it nickbuildsai.substack.com/p/…
14
I rebuilt the training process for my MNIST classifier from scratch and caught a bug: the optimizer was using the wrong model’s parameters. After several weeks on fastai’s Lesson 3, I understand how the pieces fit together and am ready to move on to NLP. nickbuildsai.substack.com/p/…
22
I finished first working versions of both fastai Chapter 4 assignments. I still need the code open beside me. nickbuildsai.substack.com/p/…
18
I could understand the training loop while reading chapter 4, but I could not reconstruct it from a blank notebook. The course teaches it bottom-up; I naturally build software top-down. nickbuildsai.substack.com/p/…
19
This week, I rebuilt gradient descent in PyTorch from memory, then tested five ResNet models on my zoysia disease classifier. ResNet-101 reached 91% validation accuracy. nickbuildsai.substack.com/p/…
32
This week, I built a regression model in Excel, then extended it into a small neural network by combining two ReLUs. nickbuildsai.substack.com/p/…
🤖 Made with AI
39
I built three fastai apps this past week nickbuildsai.substack.com/p/…
1
33