Skip to main content
Version: v0.23.x

Datasets

Overview​

The Datasets page lets you manage all datasets within your current project. From here, you can view, search, filter, and take actions on individual datasets — including launching experiments directly from the dataset view.

You can access Datasets from:

  • The top navigation tabs, or
  • The Datasets card on the homepage of a project

view-datasets

List datasets​

When you open the Datasets page, you’ll see a table showing all datasets in the current project.

Table columns​

Each row represents one dataset and includes:

  • Name
  • Number of rows
  • File size (for example, KB or MB)
  • File type (for example, CSV or Parquet)
  • Created date

You can:

  • Search by name
  • Search by file type
  • Customize columns using the column visibility toggle (top-right of the table) toggle-dataset-view

Edit dataset​

Click the checkbox icon at the top-left of the table to:

  • Select multiple datasets
  • Bulk delete selected datasets

edit-dataset

Use the Cancel button (bottom of the screen) to exit edit mode.

Row actions​

Each dataset row includes a dropdown menu with:

  • Rename — opens a rename dialog
  • Delete — removes the dataset from the project

Add new dataset​

Click the ➕ New Dataset button (top of the page) to upload your own data.

In the New Dataset screen, you can:

  • Name your dataset (optional — a default name is generated if left blank)
  • Select a Problem Type from the dropdown
  • Upload a file via drag-and-drop or file picker

Problem Type dropdown

Problem type​

The Problem Type dropdown determines how the dataset is interpreted during training and which settings are shown when you create a new experiment.

Supported problem types and dataset requirements:

Causal LM​

  • Requires an input column.
  • The output column is optional for conversation-style datasets.
  • If your dataset contains only input text (without labels), you can generate outputs using Data Generation before fine-tuning.
  • In experiment setup, this uses text fields (for example, input_column and optional output_column).

Classification LM​

  • Requires an input column and an output (label) column.
  • Output labels are validated as classification labels during training.
  • In experiment setup, this uses text-based input_column and output_column fields.

Image Classification​

  • Requires an image column containing base64-encoded or binary image data.
  • Requires an output (class-label) column.
  • Unlike object detection, this does not require a COCO annotation column.
  • In experiment setup, this uses image-specific fields such as Image Column instead of text-only input fields.

Object Detection​

  • Requires an image column and a single annotation output column.
  • Each annotation must include bbox and category fields (COCO-style structure).
  • The output column must be a single column name (not a list).
  • In experiment setup, this uses image_column and output_column (annotation) fields, not text input fields.

Multimodal Causal LM​

  • Requires an input column, an output column, and an image column.
  • The optional system prompt column is supported if provided.
  • In experiment setup, this uses both text and image fields.

Tabular​

  • Used for tabular prediction tasks (classification and regression) powered by TabH2O.
  • Requires a target column to predict.
  • Tabular datasets do not support training — predictions are run directly using a foundation model from the Predictions page.
  • Supported use cases include binary classification, multi-class classification, and regression.

Supported file formats​

  • .csv — Comma-separated values
  • .parquet or .pq — Parquet format
  • .json — JSON structured data

Once uploaded, you’ll be returned to the Datasets page and your new dataset appears in the list.

Use a demo dataset​

If you don’t have a dataset handy, you can start with one of our pre-configured demo datasets.

On the New Dataset screen, scroll down to find:

“…or get started quickly with our demo datasets”

Clicking any of the demos immediately imports that dataset into your project.

demo-datasets

Available demo datasets​

DatasetProblem TypeSizeDescription
Chatbot ConversationsCausal LM14 MB • ~12K rowsInstruction-output pairs for training conversational AI models (OASST2)
Text-to-SQLCausal LM3.7 MB • 3K rowsQuestion-SQL pairs for training models to convert natural language to SQL
Sentiment ClassificationClassification LM20 MB • ~25K rowsIMDB movie reviews for binary sentiment classification
Multi-Label Text ClassificationClassification LM5.1 MB • ~10K rowsFinancial news with multiple category labels (Reuters-21578)
note

The specific demos available may change over time — they’re designed to help you quickly test and explore LLM fine-tuning in the platform.

After importing a demo, you’ll be redirected back to the Datasets list where you can explore and use it like any other dataset.

Dataset overview page​

Clicking on a dataset opens its Overview page.

Header actions​

At the top of the page:

  • Rename the dataset via the edit icon next to the name
  • Create Experiment to kick off a fine-tuning run using this dataset
  • Delete the dataset

Metadata summary​

Right below the header, you’ll see metadata about the dataset:

  • File size
  • Row and column count
  • Token count (total tokens in the dataset)
  • Created date
  • Number of experiments launched using this dataset

Column details​

Scroll down to view the Column Details section:

  • Each column is listed with:
    • Column name
    • Max token length for that column (useful for prompt planning)
  • You can collapse or expand this section as needed

Dataset sample​

Below the column details is the Dataset Sample table:

  • Shows up to 20 preview rows
  • You can expand individual rows to view full text if it’s truncated
  • Especially helpful for verifying formatting and text content in datasets with long texts
Feedback

Related content

Data generation
Experiments