Dataset Management

This guide covers how to manage datasets in FloTorch, from creation to viewing and downloading.

Creating a Dataset

Step-by-Step Process

1. Go to your Workspace → Datasets
2. Click Create Dataset (or New Dataset)
3. In Choose Dataset Type, select one:
   • Question and Answer Pair (Q&A Pair)
   • Question and Answer Pair with Context (Q&A + Context)
4. Click Configure and enter dataset information:
   • Name: Unique identifier (lowercase letters, numbers, hyphens only)
   • Description: Optional summary of the dataset’s purpose
   • Creation Method: Options depend on selected dataset type
5. Click Create Dataset

After selecting the dataset type:

• Q&A Pair supports: Upload files, Manual creation, Auto capture, Import from HuggingFace, and Generate from PDF.
• Q&A + Context supports: Upload Q&A with context files.

Adding Content via Upload

1. Open your dataset → Add Content → Upload
2. Use drag-and-drop or click to browse
3. Select your ground truth file (required) and optionally an examples file
4. Files must be JSON or JSONL, max 10 MB each
5. Upload and confirm

The system validates format, size, and required fields.

---

Viewing Datasets

Dataset List

The Datasets section lists all datasets in your workspace with:

• Dataset name
• Description
• Dataset type (Q&A Pair or Q&A + Context)
• Creation and update timestamps

Filtering and Search

• Search: Find datasets by name or description
• Archived / Unarchived: Filter datasets by status
• Pagination: Navigate through large lists

---

Downloading Dataset Files

To download files for backup or review:

1. Open the dataset
2. In the files section, find the file you want
3. Click Download next to the file

You can use downloads to back up data, share with others, or review offline.

---

Deleting Datasets

Dataset deletion is not supported to prevent accidental data loss and protect evaluation projects that reference datasets.

If you no longer need a dataset:

• Add “DEPRECATED” or similar note in the description
• Create a new dataset with updated data instead

---

Common Issues and Solutions

File Upload Errors

“Invalid file type”
Ensure the file has a .json or .jsonl extension and contains valid JSON.

“File too large”
Reduce file size to under 10 MB. Split into multiple datasets or remove unnecessary fields if needed.

“Invalid file content”
Check that required fields are present and values match expected types.

Validation Errors

“Question cannot be empty”
All question fields must contain non-empty strings.

“Invalid JSON on line X”
For JSONL files, ensure each line is a valid JSON object.

“File must contain a JSON array”
For JSON format, wrap your data in square brackets.

---

Best Practices

File Organization

• Use clear, consistent naming for datasets
• Add version info to descriptions (e.g., “v2 - updated March 2024”)
• Download and back up important datasets regularly

Data Quality

• Validate files locally before uploading
• Start with a small sample to verify format
• Add descriptions so your team understands each dataset’s purpose

Collaboration

• Write clear descriptions for shared datasets
• Use consistent naming conventions across the team
• Notify team members when adding or changing content

---

Using Datasets in Evaluation Projects

1. Go to Evaluations → Projects → Create Project
2. Select your dataset during project setup
3. Select a Knowledge Base if your evaluation setup requires one
4. Configure and run experiments

The evaluation system uses questions from your ground truth as inputs and compares model outputs to expected answers to generate metrics.

---

Notes

• Dataset names cannot be changed after creation
• Datasets cannot be deleted
• Dataset names must start with a letter and can contain lowercase letters (a-z), numbers (0-9), and hyphens (-)
• Max 10 MB per uploaded file; 50 MB for synthetic PDF source