Also known as: evals · benchmarking · LLM evaluation
AI evaluation is the practice of systematically measuring model performance against defined tasks and criteria, using benchmarks, human review, or automated graders to quantify capability and detect regressions.
Public benchmarks suffer from contamination: once a benchmark is widely published, its contents appear in subsequent training data, and scores rise without corresponding capability gains. Held-out and continuously refreshed evaluations partially mitigate this.
For applied work, general benchmarks are weakly predictive of performance on a specific task. Teams shipping AI products almost universally end up building bespoke evaluation sets from their own data, which is unglamorous but the only reliable signal.
Hallucination
A hallucination is output from an AI model that is fluent and confident but factually incorrect or entirely fabricated, such as an invented citation, a non-existent API method, or a plausible-sounding but false statistic.
RLHF
RLHF (Reinforcement Learning from Human Feedback) is a training technique that aligns a model's behaviour with human preferences by training a reward model on human comparisons of outputs, then optimising the language model against that reward signal.
AI Red Teaming
AI red teaming is the practice of deliberately probing a model or AI-powered application for harmful, unsafe, or policy-violating behaviour before and after deployment.
Grounding
Grounding is the practice of tying a model's output to specific verifiable source material, so that claims can be traced to documents rather than resting on the model's parametric memory.
Large Language Model
A large language model (LLM) is a neural network trained on vast quantities of text to predict likely continuations of a sequence, producing systems capable of generating coherent language, answering questions, and writing code.