Convincing yet wrong content produced by an LLM, which does not check facts. Grounding lowers how often it happens but can't remove it entirely.
Also called hallucination.
Read more: Microsoft Learn
In the Ultra Transcenders books
Each book explains Fabrication in context, with comparison tables and the common traps.
Terms in this definition
- LLM
Large language model, usually a transformer network with billions of parameters that learnt to predict the next token from vast amounts of text. Compared with a small language model it is more capable across tasks, but slower and costlier.
- Grounding
Putting relevant and up-to-date facts into the prompt so that the model bases its answer on them, which cuts fabrication and makes citations possible. It is central to RAG.
Related terms
- Grounding data
Reliable content placed into the prompt when the request is made, so the model draws its answer from that material instead of guessing, which lowers fabrication.
- Hallucination
Output from a model asserting things that its sources or the facts do not support. It is gauged indirectly through groundedness, and a falling hallucination rate counts as an improvement.