Application & Purpose
Colada is Ella’s own custom BERT variant language model designed for evaluating factual accuracy. This model assesses whether one newly created sentence provides the same information as its reference.
Best Practices
- Assessing factual accuracy by evaluating whether the information conveyed in sentence A supports, denies, or is completely different from the statement in sentence B
This language model is not optimized for
- Referencing longer text blocks beyond individual sentences
- Referencing unrelated text pairs the the model is likely to categorize as “completely different”
Application & Purpose
Gaspito is a hybrid language model created by Ella which combines our own language model Maskito with OpenAI’s GPT3. In this hybrid model, GPT3’s linguistic quality is supplemented by Maskito’s factually accurate rewriting capabilities. The primary purpose of this hybrid model is to rewrite texts in a neutral style.
Best Practices
- Rewriting German and English news articles in a neutral style
- Rewriting unique texts with different wording than the original text while maintaining factual accuracy
- Simultaneous rewriting of multiple sentences in one paragraph
- Paraphrasing inadequately modified sentences whilst guaranteeing factual accuracy
Application & Purpose
Maskito is Ella’s custom sequence-to-sequence language model which is designed specifically for rewriting news texts sentence by sentence. This model utilizes a neutral style when rewriting sentences, ensuring a consistent and unbiased output.
Best Practices
- Rewriting sentences and news texts in a neutral style
Application & Purpose
GPT-3, ChatGPT and GPT-4 refer to open source large language models (LLM) within the GPT series provided by OpenAI. These decoder-only GPT models are designed for general-purpose text generation and can carry out text instructions effectively. The GPT models can also be utilized for the rapid development of prototypes.
Best Practices
- Generating fluent, coherent and well-readable texts