Glossary
AI decision-model glossary, A to Z
60 terms, each defined in plain language with links to the deeper explainers.
Swiss-hosted inference. Nothing you send is ever stored. Swiss data residency
60 pages · press / to search
A
- Decisions in productionAbstentionAbstention is when an automated system deliberately declines to decide, usually because its confidence is too low, and hands the case to a person or another system.
- Laya & the APIAct probabilityAct probability is Laya's estimate of whether an answer is safe to act on rather than escalate; its maintainers report it does not yet carry a usable signal.
- Benchmarks & datasetsAG NewsAG News is a public benchmark in which models sort news articles into four topics (World, Sports, Business, Sci/Tech), widely used to compare text classifiers.
- Calibration & scoringAutoregressive modelAn autoregressive model, such as a GPT-style chatbot, writes its output one word-piece at a time, with each new piece depending on everything it has written so far.
B
- Benchmarks & datasetsBanking77Banking77 is a public benchmark of 13,083 banking customer queries sorted into 77 closely related intents, which makes it a hard test for classifiers.
- Speed & servingBatchingBatching means processing several requests or questions together in one pass through a model, which lowers the cost per item and raises the total work done.
- Calibration & scoringBrier scoreThe Brier score rates probability forecasts by the average squared gap between the predicted probability and what actually happened; lower is better.
C
- Calibration & scoringCalibrationA model is calibrated when its confidence can be taken at face value: of all the answers it gives at 80% confidence, about 80% turn out to be right.
- Laya & the APIChoice questionA choice question asks Laya to pick one answer from a list you define, such as a department, and returns the pick plus a probability for every option.
- Models & architectureContext windowThe context window is the maximum amount of text, counted in tokens, that a model can read in one go; anything longer is cut off or must be split.
D
- Decisions in productionData residencyData residency is the requirement or promise that data is stored and processed only within a specific country or region, such as Switzerland.
- Calibration & scoringDecoder modelA decoder model reads text left to right and predicts the next word-piece, which is the design behind text-generating AI such as GPT-style chatbots.
E
- Models & architectureEncoder modelAn encoder model reads a whole text at once, looking both forwards and backwards, to understand it rather than to write new text; BERT-family models are encoders.
- Decisions in productionEntropy-based confidenceEntropy-based confidence measures how concentrated a probability distribution is on a single answer, from 1 (completely sure) to 0 (every option equally likely).
- Decisions in productionEscalationEscalation hands a case from an automated decision to a more capable handler, such as a human, a senior team or a larger model, when risk or uncertainty is high.
- Calibration & scoringExpected calibration error (ECE)Expected calibration error (ECE) is a single number for how far a model's stated confidence is from how often it is actually right, where lower is better.
F
- TrainingFew-shot learningFew-shot learning teaches a model a task from just a handful of labelled examples, either shown in the request or used for a small amount of extra training.
- TrainingFine-tuningFine-tuning is extra training of an existing model on your own labelled examples so that it specialises in your task, usually becoming more accurate.
G
- TrainingGaussian exploration noiseGaussian exploration noise is small random jitter added to a model's outputs during training so it tries slightly different answers and learns which score better.
- Speed & servingGPU inferenceGPU inference means running a trained AI model on a graphics card, whose thousands of parallel cores answer requests far faster than an ordinary processor.
H
- Decisions in productionHallucinationA hallucination is when a text-generating AI states something fluent and plausible that is not supported by its input or by fact, such as an invented category.
- Decisions in productionHuman-in-the-loopHuman-in-the-loop means people review, correct or approve some of an automated system's decisions, and their feedback is used to improve it over time.
I
- Speed & servingInference latencyInference latency is how long a model takes to answer one request, from sending the input to getting the result back, which is the wait users and agents feel.
- Decisions in productionIntent classificationIntent classification works out what a person wants from their message, such as a refund or a password reset, so it can be routed or answered correctly.
L
- Laya & the APILanguage routingLanguage routing automatically sends each request to the model best suited to its language; Laya chooses between its English and multilingual models this way.
- Laya & the APILayaLaya is an open-source (Apache-2.0) decision model by Convai Innovations that answers typed questions about text with probabilities, without generating text.
- Calibration & scoringLog scoreThe log score grades a forecast by the logarithm of the probability it gave to what actually happened, so confident wrong answers are punished very heavily.
- Models & architectureLogitsLogits are the raw scores a model gives each option before they are turned into probabilities; only the gaps between them matter, not their absolute size.
M
- Laya & the API[MASK] tokenThe [MASK] token is a special placeholder that stands in for a hidden word during training; Laya reuses it to mark the spot where each answer option is scored.
- Models & architectureMasked language model (MLM)A masked language model learns language by filling in blanks: words in a sentence are hidden and the model guesses them from the words on both sides.
- Benchmarks & datasetsMASSIVEMASSIVE is Amazon's dataset of about one million voice-assistant requests in 51 languages, used to test how well models understand intents across languages.
- Laya & the APImmBERTmmBERT is a multilingual text-understanding model in the ModernBERT style, trained on 1,800+ languages; Laya's multilingual model is built on mmBERT-base.
- Laya & the APIModernBERTModernBERT is a 2024 modernised version of the BERT text-understanding model that reads longer inputs faster; Laya's English model is built on ModernBERT-large.
N
- Benchmarks & datasetsNatural language inference (NLI)Natural language inference (NLI) is deciding whether one sentence follows from, contradicts, or is unrelated to another, and it powers many zero-shot classifiers.
- Decisions in productionnFADP (Swiss Federal Act on Data Protection)The nFADP is Switzerland's revised Federal Act on Data Protection, in force since 1 September 2023, which sets the rules for handling personal data about people.
- Models & architectureNon-autoregressive modelA non-autoregressive model produces its whole answer in one step, instead of writing it out word by word the way a chatbot does, which makes it much faster.
- Laya & the APINoul questionA noul question is a yes/no check: Laya returns the probability that a statement about your text is true, such as 'the customer asks for a refund'.
P
- Speed & servingp50 / p95 latencyp50 and p95 latency are the response times that half and 95% of requests beat, describing the typical wait and the slow tail that users actually notice.
- Calibration & scoringProper scoring ruleA proper scoring rule grades probability forecasts so that the only way to get the best expected score is to report what you honestly believe.
R
- Calibration & scoringRanked probability score (RPS)The ranked probability score (RPS) grades forecasts on ordered scales, so predicting 'high' when the truth is 'medium' costs less than predicting 'low'.
- Laya & the APIReinforcement learningReinforcement learning trains a model by trial and reward rather than by copying labelled answers; Laya uses it to reward honest probabilities.
- Calibration & scoringReliability diagramA reliability diagram is a chart of a model's stated confidence against how often it was actually right, where an honest model follows the diagonal line.
- TrainingReward shapingReward shaping adjusts or combines the rewards used in reinforcement learning to steer a model toward the behaviour you want and help it learn faster.
- Calibration & scoringRLCDRLCD (Reinforcement Learning for Calibrated Decisions) is how Laya is trained: it is rewarded for honest probabilities, so overstating confidence never pays.
S
- Laya & the APIScore questionA score question asks Laya to place something on an ordered scale you describe, such as low, medium or high urgency, and returns the expected level.
- Laya & the APIScript detectionScript detection works out which writing system a text uses, such as Latin, Cyrillic, Devanagari or Chinese characters; Laya uses it to pick the right model.
- Models & architectureSoftmaxSoftmax is the step that turns a model's raw scores for each option into probabilities that are all positive and add up to 100%.
- Calibration & scoringSpherical scoreThe spherical score is a bounded, honesty-rewarding way to grade probability forecasts, so one confident miss cannot wreck the total as it can with the log score.
- Decisions in productionSystem 1System 1 is fast, automatic, intuitive thinking; in AI the term describes models that decide in one quick pass instead of reasoning step by step.
- Decisions in productionSystem 2System 2 is slow, deliberate reasoning; in AI the term describes models that spend extra time and compute, such as writing out steps, before they answer.
T
- Calibration & scoringTemperature scalingTemperature scaling is a one-number fix applied after training that makes a model's probabilities less over-confident, without changing which answer it picks.
- Speed & servingThroughputThroughput is how much work a model server finishes per second, such as questions answered, and depends on the hardware, batching and length of the inputs.
- Calibration & scoringTokenizerA tokenizer chops text into the small word-pieces (tokens) a model actually reads, and its vocabulary decides which languages and scripts the model handles well.
- Laya & the APITyped decisionsTyped decisions are answers limited to a format you declare up front (a pick from a list, a level on a scale, or yes/no) and returned with probabilities.
Z
- Decisions in productionZero data retentionZero data retention means a service processes what you send and then discards it, never saving the content to a database, log or training set.
- Models & architectureZero-shot classificationZero-shot means a model handles a task it was never specifically trained for, using only the labels or descriptions you give it at the time of the request.
Get started
Try it on your own text
Send any text with your own typed questions and get a calibrated probability for every answer. 6k free tokens, no card needed; then $0.0357 per 1M input tokens — 15% below Jev.
Swiss-hosted inference. Nothing you send is ever stored.