Skip to content

Glossary

The words the app and this guide use, in alphabetical order. For the terms of the Python libraries, see the ML4 glossary and the MLI glossary.

A

Activity. The list of what happened and what is running, in one workspace or across all of them. Its Needs you view collects what waits for you. See Workspaces and sessions.

Agent. A coding agent, such as Codex, Gemini CLI, Claude Code or Cursor, that ParleML runs on a task. It works only through ParleML's tools, within what its profile and token allow. See Settings and agents.

Alerts. A dataset's written assessment: its shape, quality problems and suspicious columns. See Datasets and columns.

Archetype. A group of rows the model treats alike, described by its size, its outcome rate and the features that define it. It describes the model, not a customer segment by itself. See Results and evidence.

B

Base time. A column flag that marks the column holding each row's time. It anchors "days since" features, forecasts and chronological splits. See Datasets and columns.

Baseline. The score of a model that ignores every feature and predicts the average or the class balance. A model is worth using only by how far it beats this floor. See Results and evidence.

Batch & Diagnose. The experiment view that scores a whole dataset. With the true outcome given, it also measures the model on those rows. See Score a new file.

Builder token. An API token that can build inside the workspaces it is bound to: run experiments, explain, score and register models. It expires within 90 days and can never promote or delete. See Access, permissions and data.

C

Calibration. A correction that makes a model's probabilities honest: cases scored 0.3 turn out positive about 30% of the time. See Results and evidence.

Catalog. The index of everything in the data root: workspaces, datasets, experiments, models and knowledge bases. It can be rebuilt from the files. See Configuration.

Challenge. An independent, sceptical review of a candidate model's evidence before release, run by the Challenge the candidate skill. A release record states whether it passed. See Models and release.

Chronological split. A train/test split by time: the test rows come after the training rows. Use it when the model will predict the future. See Experiments.

Compare. The side-by-side view of two to four experiments, models or datasets. See Compare.

Confidence interval. The range a score probably lies in, given how many test rows there are. A wide interval means the score is imprecise. See Results and evidence.

Connection. A saved link to an outside system: a database, MLflow, a tool server or an embedding provider. Its secret is kept in the vault. See Settings and agents.

D

Data root. The one directory where ParleML keeps all its state. In the Docker image it is /app/data_root, in the parleml-data volume. Back it up as a whole. See Install and operate ParleML.

Dataset. A table in a workspace, with column roles, quality checks and exploration views. See Datasets and columns.

Decision. A question an agent asks a person, and the answer. Only people answer decisions. The workspace's Decisions page lists them. See Workspaces and sessions.

Decision threshold. The score above which a case counts as positive. ParleML tunes it at training, and you can test others under Decision & use. See Results and evidence.

Diagnostics. The read-only Settings page that shows this installation: versions, hardware, job slots, sign-in mode and the models on the server. See Configuration.

Documents. A workspace's written records, such as reports and model cards an agent or a person published. See Workspaces and sessions.

Drift. A change between the data a model was trained on and the data it scores now. Scores on drifted data are less reliable. See Datasets and columns.

E

EDA. Exploratory data analysis: a dataset's profile of distributions, missing values and associations. It fills the Charts, Alerts and Leakage tabs. See Datasets and columns.

Evaluation metric. The score an experiment optimises and reports, such as ROC-AUC or log loss. Each kind of problem has a default. See Experiments.

Experiment. One configured, measured attempt to model a target on a dataset. One execution of it is a run. See Experiments.

Explanations. What describes why a model behaves as it does: drivers, archetypes, rules and diagnostics. You compute them with Compute explanations and read them under Interpretation. See Results and evidence.

F

Fairness attribute. A column flagged so results are broken down by its groups. It is independent of Ignore, so a sensitive column can be checked without the model using it. See Datasets and columns.

Foundation model. A model pre-trained on many tables that predicts on yours without being trained on it. ParleML uses one for the signal check and as an optional ensemble member. See Datasets and columns.

G

Grouped split. A split that keeps related rows, such as one customer's, on the same side of train and test. See Experiments.

H

Held-out rows. Rows never seen in training: the test set. The headline score, fairness, permutation importance, the error tree and the residual surrogate are computed on them when the experiment has them. See Results and evidence.

I

Intended use. The statement of what a workspace's model is approved for, and what it is not. A release is judged against it. See Models and release.

Investigations. A workspace's agent sessions: their questions, findings, reports and history. See Workspaces and sessions.

K

Knowledge base. The tables a workspace can read, with the joins between them and the recipes that build training datasets. See Connect a database.

L

Leaderboard. The models an experiment tried, ranked by their validation score. See Results and evidence.

Leakage. Information in a column that would not exist when the prediction is made, so the model looks better than it will be. Also the dataset tab that lists suspected leaks. See Datasets and columns.

Lifecycle stage. Where a model is in its life: Trained, Validated, Promoted, Shadow, Production or Archived. Only a person moves a model between stages. See Models and release.

M

MCP. The Model Context Protocol, the standard agents use to call ParleML's tools. ParleML serves it at /mcp. See the MCP tools.

MLflow. An open-source tracker for runs and models. ParleML can log supervised runs to it, so the fitted model is a standard MLflow model. See Models and release.

Model. A fitted result of an experiment, registered with a name, risk level and owner, so it can be reviewed, staged and used. See Models and release.

Model card. The generated report on an experiment or model: data, features, importance, model choice and limitations. It is on the Report tab. See Results and evidence.

N

Noise reference. A check that ranks every feature against columns of pure noise, with three quick fits before training. Features that rank no better than noise are listed under the experiment's Run settings; nothing is dropped. See Experiments.

O

Open mode. A server with sign-in off: anyone who can reach it can use it, with administrator rights. It is the default. See Set up single sign-on.

Operating point. The threshold at which a model is to be used, and how it was chosen, for example to fit a review team's capacity. A release records it as applied or documented only. See Models and release.

P

Positive class. The outcome counted as positive for metrics and the threshold. It defaults to the rarer class. See Experiments.

Prediction moment. The point in time a prediction would be made. A recipe's point-in-time check flags any column that reads data from after it. See Connect a database.

Preflight. The check that a profile's agents are signed in and its connections answer. See Add your own skill pack.

Profile. The rules a workspace's sessions follow: which sessions people can start, which agent and models, skill packs, auto-approval, limits and connections. Standard is the default. See Settings and agents.

R

Reason codes. A per-row file with each case's prediction, its top drivers and the rules it matched. See Score a new file.

Recipe. A saved SQL query in a knowledge base that turns source tables into a training dataset, checked for leakage against its prediction moment. See Connect a database.

Release record. The written result of a release review: approval, challenge result, operating point, conditions and limitations. It is separate from the lifecycle stage. See Models and release.

Risk level. A workspace's or model's label: Low, Medium, High or Regulatory. See Models and release.

Rule. A readable if-then statement that describes a pattern, with its numbers on training and test rows. Business Rules come from a separate compact model on the original columns. See Results and evidence.

Run. One execution of an experiment. See Experiments.

S

Scope. What an API token may do: analyst (read), compute (read plus bounded calculations) or builder (build in named workspaces). See API reference.

Scoring package. A model downloaded as a zip with a pinned requirements.txt and a ready-to-run score.py, for predicting without a ParleML server. See Use your model in Python.

Session. An agent working on a task, with its steps, questions and outcome. Sessions are numbered in the order they start, such as S-82. See Workspaces and sessions.

Signal check. A quick test, with a foundation model and no training, of whether the data holds anything to learn about the target. A good result is a lower bound. See Datasets and columns.

Skill. One procedure an agent follows in a session, such as framing a problem or challenging a candidate model. Skills drive ParleML's tools; they don't write modelling code. See Add your own skill pack.

Skill pack. A set of skills, templates and rules. A profile selects the packs its sessions use. See Add your own skill pack.

Surrogate model. A simple model fitted to imitate a complex one so it can be read. It is an approximation, with its own accuracy. See Results and evidence.

T

Target. The column a supervised experiment predicts. See Your first model.

Task. The kind of work a session starts with, such as delivering a model or checking one in operation. A profile's Startable sessions set which ones people can choose. See Workspaces and sessions.

Template. A ready-made training recipe, such as Quick Explore or Default, chosen in the launch wizard. See Experiments.

Token. A credential for scripts and agents, shown once when minted, with a scope. See API reference.

V

Validation score. The score on rows set aside to choose between models. Because choosing can favour luck, trust the held-out score as well. See Results and evidence.

Vault. The encrypted store of connection secrets. Its key lives in the data root. See Settings and agents.

W

Workspace. One business problem, with its data, experiments, models, sessions, documents and decisions. See Workspaces and sessions.

Workspace data catalog. A knowledge base made from the workspace's own datasets and published results, so you can query and join them with SQL. See API reference.

Next