LLM Experiments
An open, ongoing notebook of small, sharp experiments on transformer internals — attention patterns, fine-tuning behavior, and failure modes.
Problem
Understanding *why* a fine-tuned model behaves the way it does is usually harder than getting it to a working state in the first place.
Approach
A running series of small, reproducible experiments — each isolating one variable (learning rate schedule, adapter rank, data mixture) and documenting the result, whether positive or not.
Architecture
Each experiment ships as a standalone, reproducible notebook with pinned dependencies, tracked runs, and a short written conclusion — designed to be read individually, not as one monolithic project.
Results
- 30+ documented experiments, several of which fed directly into production fine-tuning decisions.
- Became a reference she returns to before starting any new fine-tuning effort.
Future work
Turning the strongest experiments into a short written series on practical fine-tuning intuition.