#Working with Datasets

eval_752 gives you three ways to get a dataset into the system:

1. Build one in the browser — the Dataset Builder GUI
2. Import from Hugging Face — guided wizard with column mapping
3. Upload an .eval752.zip — for pre-packaged datasets

Pick whichever fits your situation:

#Building a dataset in the browser

Use the Dataset Builder when your benchmark doesn't exist anywhere yet — it lives in your head, a document, or scattered notes.

1. Open Datasets → Dataset Builder → Open Builder
2. Step 1: Name, intent, and description
3. Step 2: Create sections (logical groupings for your items)
4. Step 3: Add items — type in prompts and answers, duplicate existing items, drag items between sections, attach local images
5. Step 4: Review counts and composition
6. Click Publish dataset

Publishing creates a formal dataset you can use in Runs. The draft stays available for future edits.

Attached images appear as real thumbnails while you edit. During runs, image assets are sent to the model if the provider supports vision.

#Importing from Hugging Face

Use this when the dataset already exists on HF.

1. Open Datasets → Start import wizard
2. Enter the dataset path, split, and preview limit
3. Optionally set a display name
4. Click Preview dataset
5. Map the detected columns:
   • Prompt column (required)
   • Answer column (required)
   • Choices, type, metadata columns (optional)
6. Review and import

The wizard stores a normalized copy in your local database. After import, the dataset appears immediately in Runs.

#Uploading an .eval752.zip

For pre-packaged datasets:

1. Open Datasets → Upload eval752 package
2. Pick the .eval752.zip file
3. Optionally override the display name
4. Click Upload dataset

The package gets unpacked into dataset records. Embedded assets (images) are restored so they continue to work in runs.

#Exploring and exporting

Each dataset card in the main list supports:

• View — opens the explorer dialog with search, section filters, and pagination
• Export — downloads the full dataset as .eval752.zip
• Section chips — shows how items are grouped

Inside the explorer you can:

• Search items by text
• Filter by section
• Preview attached images
• Export a filtered subset as .eval752.zip

This is the best way to verify a dataset before running an evaluation.

#What's not supported yet

• Direct CSV / JSON / JSONL file upload from the browser
• Direct .llmset.zip upload

If you have datasets in these formats, convert them to .eval752.zip first using the project's packaging tools.

#Tips

• Use meaningful names. Dataset names show up everywhere — in Runs, Comparison, and exports. "MMLU Math Subset" is better than "Dataset 1."
• Use section filters to create smaller review bundles before sharing.
• If a Hugging Face dataset needs extra auth or configuration not exposed by the wizard, use a simpler public dataset for now.