Skip to main content
Datasets are the training data for your fine-tuning jobs. You upload a dataset once and reference it from any number of jobs.
Valkyrie Data screen

The Data screen: each file with its format check and what it taught.

​
Upload a dataset

1

Open Data

In the dashboard, go to Data and click Upload data.
2

Upload your file

Drop your file or click Choose file (JSONL, JSON, CSV, Parquet, or text, up to 2 GB), and optionally give it a name. Valkyrie stores the file for you and checks its format as soon as it lands.
3

Check the format

The dataset’s page opens with the Format check result, the detected format, and the columns it found. Fix the file and upload it again if the check reports Needs fixing.
4

Train on it

Click Fine-tune with these to start a fine-tune with this file selected, or pick it under Your examples in New fine-tune.
See Data for everything on these screens.
Data not in training format yet? Attach your spreadsheet (Excel, CSV, JSON, Parquet) or text to the AI assistant and say what you need. It converts every row into training examples and saves them as a new dataset.

​
Formatting your data

Fine-tuning expects structured conversational or instruction data. A common, portable shape is JSON Lines (.jsonl), one training example per line, for example a chat-style record:
Instruction and response pairs, prompt and completion, plain text, and label columns for classification are accepted too. No data yet? The upload screen offers a ready-made sample, customer-sentiment.jsonl or customer-sentiment.csv, to try the whole flow.
Keep examples consistent in structure and representative of how the model will be prompted in production. Quality and consistency matter more than raw volume.

​
Managing datasets

From the Data screen (or via the MCP tools) you can:
  • List your datasets, check their format, and see which model versions each one taught.
  • Download a dataset with a temporary signed URL.
  • Delete a dataset you no longer need.
Datasets are stored in object storage. Deleting a dataset does not affect models you’ve already trained from it.