🧐 Choose a dataset type#
FAQs#
What? Different datasets?
First things first, Argilla offers two generations of datasets. The new FeedbackDataset and the older datasets called DatasetForTextClassification, DatasetForTokenClassification, and DatasetForText2Text.
Why the new FeedbackDataset?
In short, the FeedbackDataset is a fully configurable dataset that can be used for any NLP task including LLM-focused tasks. The older datasets are focused on a single NLP task. As a result, the FeedbackDataset is more flexible and can be used for a wider range of use cases including all NLP tasks of the older datasets. The older datasets are more feature-rich in certain points but no new features are introduced, on the other hand, the FeedbackDataset is currently less feature-rich in certain points but new features will actively be added over time.
Will the older datasets be deprecated?
We will continue to maintain the older datasets for the foreseeable future, but we recommend using the new FeedbackDataset, which is going to be the core of Argilla 2.0.
When should I use older datasets?
At the moment, the older datasets are better when doing basic Text Classification or Token Classification. They provide full support for metadata-filters, bulk-annotation, weak supervision, active learning and vector search.
When should I use FeedbackDataset better?
The FeedbackDataset is better when you need to do more complex tasks that need to be represented in one coherent UI. This is extremely useful for LLM workflows where you need to do multiple tasks on the same record. The FeedbackDataset also supports multiple annotators per record, customizable tasks and synchronization with a database. However, it does not support weak supervision or active learning yet.
When will all the cool features of the older datasets be available in the FeedbackDataset?
We are working on it! We will be adding new features to the FeedbackDataset over time. If you need a specific feature, please let us know on GitHub or Slack so we can prioritize it.
Table comparison#
NLP Tasks#
Task / Dataset |
FeedbackDataset |
Older datasets |
|---|---|---|
Text classification |
✔️ |
✔️ |
Token classification |
✔️ |
✔️ |
Summarization |
✔️ |
✔️ |
Translation |
✔️ |
✔️ |
NLI |
✔️ |
✔️ |
Sentence Similarity |
✔️ |
|
Question Answering |
✔️ |
|
RLHF (SFT) |
✔️ |
|
RLHF (RM) |
✔️ |
|
RLHF (PPO) |
✔️ |
|
RLHF (DPO) |
✔️ |
|
RAG |
✔️ |
|
Image support |
✔️ |
|
Overlapping spans |
✔️ |
|
And many more |
✔️ |
Annotation workflows#
Task / Dataset |
FeedbackDataset |
Older datasets |
|---|---|---|
bulk annotation |
✔️ |
✔️ |
vector search |
✔️ |
✔️ |
active learning |
✔️ |
|
weak supervision |
✔️ |
User and team management#
Features |
FeedbackDataset |
Older datasets |
|---|---|---|
Multiple annotators per record |
✔️ |
|
Multiple-tasks in one UI |
✔️ |
|
Synchronization with database |
✔️ |
UI Sorting and filtering and querying#
Features |
FeedbackDataset |
Older datasets |
|---|---|---|
Record status filters |
✔️ |
✔️ |
Text query |
✔️ |
✔️ |
Metadata filters |
✔️ |
✔️ |
Sorting |
✔️ |
✔️ |
Prediction filters |
✔️ |
✔️ |
Annotation filters |
✔️ |
✔️ |
Similarity search |
✔️ |
✔️ |