Stage 1 · Lesson 4 of 6
Evals and safety testing
Once a model is trained, companies need to know how good and how safe it is. An , or eval, is a set of test with a way to score the answers, sometimes by code, sometimes by people with a rubric. Human raters are used where judgement matters: is this medical answer correct, is this code secure, is this tone appropriate?
, often called red teaming, means deliberately trying to make a model misbehave, for example by coaxing it into harmful instructions or leaking private data, so the problem can be fixed before release. Some data work is exactly this: writing tricky prompts, scoring answers and documenting failures.
Common traptreating safety tasks as a game of "breaking" the model. The useful output is a clear, reproducible report of what failed and why.
A question to check yourself, an AI practice session and 2 cheat sheet entries come with the full roadmap. Unlock
Resources for this stage
Start with· follow one of these
- CoursePaidGrokking Modern AI Fundamentals
Design Gurus · Beginner · $72 one-time
A structured course on how modern AI models are built and trained, for the full picture behind your tasks
- VideoFreeDeep Dive into Large Language Models like ChatGPT
Andrej Karpathy · Intermediate
Free: Karpathy walks through pre-training, fine-tuning and RLHF, including the role of human labellers