Upload an SFT dataset¶

SFT datasets are the Language type with Prompt and Target columns. The target is the completion the model should copy.

Prerequisites¶

An Advanced Training project.
A CSV or Parquet file (or Hugging Face parquet folder) for SFT data. See SFT datasets.

2. New dataset¶

Click New dataset. Choose category Language, then type SFT.

Create dataset modal with SFT type — Creating a new SFT dataset.¶

3. Upload files¶

Upload a local file or import from Hugging Face.

SFT dataset file upload — Uploading SFT dataset files.¶

4. Map columns¶

On Data, map columns to Prompt and Target. Save both before you leave the page or continue in the wizard.

SFT dataset column mapping step — Selecting prompt and target columns.¶

5. Preview¶

Check a sample of rows after mapping.

Dataset preview table — Preview of uploaded SFT rows.¶

6. Saved dataset¶

Save. The dataset shows up in the list and in the experiment wizard.

Datasets list with new SFT dataset — The dataset listed after save.¶

From the terminal¶

Create the dataset.

Install the CLI and sign in (Authentication):

pip install "agilerl[arena]"
arena login
# or: export ARENA_API_KEY=arena_pat_…

arena datasets create support-tickets-sft --category sft --column-mapping '{"prompt":"instruction","target":"response"}' --file ./sft.csv

category must be sft. If your file already uses prompt and target, put those names on both sides of the mapping.

Python¶

from agilerl.arena import ArenaClient

with ArenaClient() as client:
    client.create_dataset(
        name="support-tickets-sft",
        category="sft",
        file="./sft.csv",
        column_mapping={"prompt": "instruction", "target": "response"},
    )