The Alchemist’s Workshop
Crafting ideas into reality: models from scratch, experiments and research notes.
total 4
ElasticSkip: Where You Cut Matters More Than How You Heal the Cut
Compressing LLM query encoders for dense retrieval: skip the middle, keep the tail, and serve every depth from one base model.
$ cat elasticskip →
MOGIC: Teaching a Small Model to Think Like an Oracle
Our ICML 2025 paper: a slow, all-knowing oracle trains a fast disciple for extreme classification, for 1–2% more accuracy at zero extra inference cost.
$ cat mogic →
GPT from scratch
Language models have become increasingly popular in recent times. People are using it for various purposes, including information retrieval, language assistance and code…
$ cat scratchGPT →
Bewakoof Detector
I trained this cool image classifier using the fastai library to mess with my friend who we both agree is bewakoof with directions.
$ cat bewakoof-detector →No matching items