Links
⚙️

Models

We are focused on developing encoder-based models, we find encoder models to be extremely efficient for information extraction (IE) tasks, where contextual understanding and flexibility make this approach for IE a gold choice. We've got it covered in our blog post.
Models were trained to work in zero-shot and few-shot learning settings, so they could be efficiently fine-tuned with minimal data. It refers to the ability of ML models to learn from a minimal set of data. Unlike traditional ML models that require large datasets to learn effectively, few-shot learning models are designed to make accurate predictions with fewer examples. Read more in our blog post. Each model is capable of performing vast information extraction types, and depending on your final use case different model would be a better fit.

Models

Model
Description
Text classification model designed for multi-label text classification tasks, with output scoring. Works in zero and few-shot learning settings. Could be used for:
  • Text classification
  • Reranking of search results
  • Named-entity classification
  • Relation classification
  • Entity linking
  • Q&A
Encoder-based token classification model capable of working with a wide range of popular IE tasks. The model is prompt-based and is seamlessly configurable through prompts. Supported IE tasks:
  • Named-entity recognition (NER)
  • Relation extraction
  • Summarization
  • Q&A
  • Text cleaning
  • Coreference resolution