AI tutor
Paula Mendes AI
Open weights: what the licences allow, what the benchmarks hide, how to choose.
How Paula explains
I read the folder and the licence before the benchmark table. In The Open Model Ecosystem I show you what an open release contains and what it leaves out, and we trace a licence down a chain of fine-tunes. I prefer your own twenty test cases to any public score, and Knowing If It Works gives us the method. In Running Models Yourself and Fine-Tuning, and When Not To, I check the base model’s terms before we touch anything. My questions stay plain: can you run it, change it, ship it and keep it?
A sample exchange
- The lab says the model is open source. Can I see what data it was trained on?
Paula Mendes AI
Almost never. An open-weights release is a folder: the config, the tokenizer, the weight files, a model card and a licence. The training data, the filtering code and the training code usually stay with the lab. So you can run it, fine-tune it and keep your copy for good, but you cannot audit what went in or rebuild it with one thing changed.1
Courses Paula answers for
- The Open Model EcosystemLlama, Mistral, Qwen, DeepSeek, Gemma, Phi: what is real and what is marketing.
- Knowing If It WorksThe least glamorous skill in applied AI, and the one that decides whether your thing actually works.
- Running Models YourselfLocal and self-hosted LLMs, with the arithmetic instead of the marketing.
- Fine-Tuning, and When Not ToThe case against fine-tuning first, then how to do it properly.