An Albanian Edge LLM - Part 1: The Vision
Albanian is a low-resource language with around 7.5M speakers.
ReadExperiments
I built it to find out.
Working notes from things I tested, broke, rebuilt, and eventually stopped.
Experiments
Builds, tests, failures, and whatever survived contact with reality.
An Albanian Edge LLM - Part 2: Sourcing the Data
I ended up building 22,369 training pairs from scratch using multiple AI models.
ReadBuilding an Albanian Edge LLM - Part 3: The Roast
I tried a bunch of different approaches. Most failed!
ReadBuilding an Albanian LLM - Part 4: The Pour: Quantization & iOS Deployment
I got a 7GB model down to 934MB using 4-bit quantization. Then I used Apple's MLX framework to run it on iPhone.
ReadBuilding an Albanian AI - Part 5: Results and Future Plans
69.2% accuracy on my tiny 13-phrase test set. Basic phrases work. Proverbs mostly fail.
Read