MakataResearch & Learning
Understanding
what we’re building.
A growing body of study notes on the mathematics, models, and methods behind language AI. Start with an idea. Work through it. See it move.
Explore the libraryEight connected topicsFrom mathematical foundations to a defensible experiment.
Learn by changing thingsWorked examples, equations, and interactive SVG explanations.
Evidence stays visiblePrimary sources and explicit assumptions throughout.
01
Foundations
Build the language of learning, then turn it into code.
Mathematics for Machine Learning
From a vector on a plane to a gradient through a model: build the mathematical language needed to understand what learning actually computes.
Read, calculate, explore
Neural Networks & Deep Learning
Understand a neural network as a composition of parameterized functions. Follow real values forward, gradients backward, and the evidence needed to call a model useful.
Read, calculate, explore
PyTorch & Practical Model Training
Turn the mathematics into an inspectable training loop. Learn how tensors, autograd, modules, batches, and checkpoints fit together in a reproducible CPU example.
Read, calculate, explore
02
Models
Understand the architecture and what adaptation changes.
Transformers & LLM Architecture
Follow text through tokenization, embeddings, attention, and a decoder block. See how a distribution over the next token becomes a generated sequence.
Read, calculate, explore
LLM Training & Fine-Tuning
Compare pretraining, continued pretraining, supervised adaptation, and parameter-efficient methods. Make model provenance and experimental tradeoffs explicit before using compute.
Read, calculate, explore
03
Research practice
Make data, evidence, and computing choices you can explain.
Dataset Engineering for LLMs
Treat a dataset as a documented set of decisions. Follow language data from collection and rights review through cleaning, partitioning, tokenization, and contamination checks.
Read, calculate, explore
LLM Evaluation & Research Methodology
Ask what a score measures, how reliable it is, and what changed. Build an evaluation argument from baselines, controlled comparisons, uncertainty, and Philippine-language context.
Read, calculate, explore
GPU Computing & Training Infrastructure
Understand where computation and memory go during training. Estimate costs with stated assumptions, then reason about precision, batching, checkpointing, and distributed execution.
Read, calculate, explore
Original Makata research
From learning the methods to documenting the evidence.
Report 001 documents our Tagalog comprehension baselines. Report 002 follows dataset screening, QLoRA training, and paired evaluation: an observed increase whose uncertainty includes no improvement. Both include methods, evidence, and limitations.
Study notes, with a research purpose.
These articles explain established ideas and external research as we build Makata’s foundations. Interactive examples are small teaching models, not benchmark results. Original measurements from Makata’s internal experimental adapter belong in Research Studies. Its weights remain private, and a reliable comprehension gain is not established.