Key Concepts
These are the core concepts you will encounter throughout DashAI. Each term corresponds to something you can see and interact with directly in the interface.
Dataset
A dataset is the collection of data you work with. It is typically a structured table (CSV, Excel, JSON). Every workflow in DashAI begins by loading a dataset into the Datasets module.
Once loaded, a dataset is persistent — it stays in DashAI until you delete it, and it can be used across multiple notebooks and sessions.
DashAI provides dataloaders for common formats (CSV, Excel, JSON). You can create custom dataloaders to support additional data sources and install them via the Plugins module.
Notebook
A notebook is a non-destructive workspace attached to a dataset. It lets you apply a sequence of Explorers and Converters to a working copy of the data, while the original dataset remains unchanged.
Each operation you add appears in the notebook timeline, so your transformation process is always visible and auditable. When the result is ready, you save it as a new dataset with Save as New Dataset.
Explorer
An explorer is a visualization or analysis tool in the notebook's EXPLORE tab. Explorers read the current state of the data and produce a result — a chart, a statistical table, a heatmap — without modifying the data.
DashAI comes with explorers organized into five categories: Preview Inspection, Relationship Analysis, Statistical Analysis, Distribution Analysis, and Multidimensional Analysis.
You can also develop custom explorers and install additional ones through the Plugins module.
Converter
A converter is a data transformation tool in the notebook's CONVERT tab. Converters modify the data — encoding categories, scaling numerical values, removing missing values, reducing dimensionality, and more. Each converter is applied to a specific set of columns and rows, and the effect is immediately visible in the dataset preview.
DashAI includes more than 30 converters organized into eight categories: Basic Preprocessing, Encoding, Scaling and Normalization, Dimensionality Reduction, Feature Selection, Polynomial & Kernel Methods, Resampling & Class Balancing, and Advanced Preprocessing.
You can create custom converters and install additional ones from the Plugins module.
Task
A task defines the type of ML problem you want to solve. The task determines which models are available, which column types are valid as inputs and outputs, and which metrics are used to evaluate results.
DashAI's Models module supports four core tasks:
| Task | What it predicts |
|---|---|
| Tabular Classification | A category from structured table data |
| Regression | A continuous numerical value from structured table data |
| Text Classification | A category assigned to a text document |
| Translation | Text converted from one language to another |
The Generative module supports separate tasks for text and image generation. See Generative AI for details.
You can extend DashAI by developing new tasks and installing them via the Plugins module.
Session
A session is the central unit of work in the Models module. It ties together a dataset, a task, a column configuration (which columns are inputs, which is the target), and a data split strategy. All models you train for that problem live inside the same session, making it easy to compare them.
Sessions are grouped by task type in the left sidebar. You can have multiple sessions for the same dataset — for example, to explore different input column combinations or split strategies.
Model
A model is a specific algorithm added to a session. Each model has its own name, hyperparameter configuration, and training status. You can add multiple models of the same or different types to the same session and train them independently or all at once with Run All.
DashAI includes a growing library of models powered by scikit-learn, Hugging Face, PyTorch, and TensorFlow.
You can develop new model integrations and discover additional models through the Plugins module.
Training Status
Each model in a session has one of five statuses:
| Status | Meaning |
|---|---|
| Not Started | The model has been added but never trained |
| Delivered | The models has been queued |
| Started | The training process has started |
| Finished | Training completed successfully |
| Error | Training failed — review parameters or data |
Metric
Metrics measure how well a trained model performs on each data split. DashAI automatically calculates metrics appropriate for the type of task you are solving, allowing you to understand your model's effectiveness across different data splits.
For classification tasks, you can review metrics like Accuracy (percentage of correct predictions), F1 (balance between precision and recall), Precision and Recall (true positives relative to total positives and total actual positives), or ROCAUC (model's ability to distinguish between classes). Additional metrics include LogLoss, HammingDistance, and CohenKappa for deeper analysis.
For regression tasks, metrics focus on prediction error: RMSE (root mean squared error) penalizes larger errors, while MAE (mean absolute error) provides an average error magnitude.
For translation tasks, metrics include BLEU (similarity to reference translations) and TER (number of edits needed to match a reference).
DashAI includes a core set of metrics appropriate for each task. If you need additional metrics for specialized use cases, you can develop custom metrics and add them via the Plugins module.
Prediction
A prediction is the output of a trained model applied to new data. DashAI supports two prediction modes: Dataset Predictions (run the model over an entire loaded dataset) and Manual Predictions (enter specific values row by row in the interface).
Explainer
An explainer is a tool for interpreting a trained model's behavior. Global explainers analyze model behavior across the full dataset (e.g., which features matter most overall). Local explainers analyze a specific prediction (e.g., why did the model produce this output for this particular record).
DashAI provides built-in explainers including Kernel SHAP, Permutation Feature Importance, and Partial Dependence Plots.
You can create custom explainers and install additional interpretability tools from the Plugins module.
Job
Long-running operations — training a model, running an explorer, generating a prediction — are executed as background jobs. The Job Queue at the bottom right of the screen shows active and completed jobs so you can continue working while an operation runs.