Room 1
The Prompt Atrium
A visitor starts with instructions, constraints, examples, and refusal boundaries. The atrium shows how small wording changes alter the path a model takes through a task.
Public concept museum for language models
Walk through LLM ideas as exhibits, not jargon.
LLMopedia Observatory turns abstract model behavior into rooms a reader can inspect: prompt signals, context windows, evaluation habits, retrieval trails, alignment boundaries, and the social language that surrounds AI systems. The site is written for curious operators, editors, builders, students, and policy readers who need sturdy mental models before they repeat a new term.
Today's viewing rule: if a concept cannot be explained without a model name, it is not ready for the wall.
Floor guide
How the observatory reads a model concept
Room 2
The Context Conservatory
This room treats retrieved passages, chat history, tool outputs, and memory as fragile specimens. It explains why relevant context helps and why extra context can blur the answer.
Room 3
The Evaluation Balcony
Instead of announcing that a model is good or bad, the balcony separates scoring rubrics, human review, adversarial tests, and production feedback into visible layers.
Lens
Clarifies what a term means before comparing claims about it.
Scale
Weighs reliability, traceability, and user cost instead of hype.
Map
Places concepts near the workflows where people actually meet them.
Cabinet
Keeps short definitions beside longer explanations for repeat visits.
Editorial method
A concept is useful when it can survive a guided tour.
The observatory avoids treating LLM knowledge as a pile of entries. A strong explainer starts with the visitor's question, names the concept plainly, shows where the concept appears in a real workflow, and records what evidence would change the explanation. That rhythm keeps the site close to how AI systems are used in practice: a model answers, a person judges, a source anchors the claim, and a tool or policy changes the next step. LLMopedia's home floor is therefore less like a shelf and more like a working exhibition: readers can compare rooms, instruments, and viewing notes before following a single article into the deeper collection.
What belongs here
Exhibits for terms people actually rely on.
The collection favors concepts that affect decisions: hallucination, grounding, token budgets, tool calling, retrieval, system messages, red teaming, eval drift, guardrails, model cards, provenance, and answer engine visibility. Each subject is handled as a public object with a label, a room, a practical example, and a caution about overclaiming. That makes the site useful to a reader who wants a quick definition and to a team that needs shared language for reviews, prompts, procurement, or documentation.