The Data Loader, as far as I can tell, does not support Custom Metadata (and we've been given no indication if this will be available in the future). In fact, custom metadata isn't "real" data, and is actually documented in the Metadata API documentation. You could choose to use Ant and build the appropriate files that way, or you can follow the other instructions for simply loading this metadata.
In the implementation manual, the guide tells you how to import custom metadata:
Edit: The older version, included below in the original quote, has been deprecated. Please use this version instead. It also now includes a Deploy to Salesforce button to make it easier to deploy in fewer steps.
Use the custom metadata loader to bulk load records to your custom
metadata types from a .csv file. The custom metadata loader lets you
load up to 200 records with a single call.
Download the tool from GitHub. Deploy the package to your org via Workbench. Note that you have to create a .zip file of the contents of
the custom_md_loader directory instead of zipping up the directory
- Create a .csv file with a header that contains the custom metadata type’s field API names. Either the Label field or the Developer Name
field is required. See sample.csv in your download for an example. If
your org is namespaced, be sure to include the namespace prefix in
- From Setup, assign the Custom Metadata Loader permission set to the appropriate users, including yourself.
- Select Custom Metadata Loader from the App Picker.
- Go to the Custom Metadata Loader tab. The app prompts you to configure your Remote Site Settings if you haven’t already done so.
- Select your .csv file and the corresponding custom metadata type.
- Click Create custom metadata to bulk load the records from the .csv file.