google/gemma-2b-aps-it
Gemma-2b-aps-it is a 2.6 billion parameter generative model developed by Google, specifically designed for abstractive proposition segmentation (APS) or claim extraction. This model segments text into individual facts and restates them as full sentences, making it a specialized research tool. It is optimized for tasks requiring the breakdown of text content into meaningful components, such as grounding, retrieval, fact-checking, and evaluation of generation tasks. The model was trained with a context length of 8192 tokens.
Loading preview...
Overview
Gemma-2b-aps-it is a 2.6 billion parameter generative model from Google, specifically engineered for abstractive proposition segmentation (APS), also known as claim extraction. This model takes a text passage and identifies individual facts, statements, or ideas, then restates them as distinct, full sentences with minor modifications from the original text. It is a research tool designed to break down complex text into its constituent propositions.
Key Capabilities
- Abstractive Proposition Segmentation: Accurately segments text into individual, restated propositions.
- Research Tool: Ideal for applications like grounding, retrieval, fact-checking, and evaluating generative AI outputs by dissecting claims.
- English Text Only: Specialized for processing English language content.
- Context Length: Supports an 8192-token context window.
Training and Limitations
The model was trained on synthetically generated examples, where input passages were created using Gemini Ultra and propositions were generated by a teacher LLM (Gemini Pro) trained on a filtered ROSE dataset. It is important to note that Gemma-2b-aps-it is only suitable for abstractive proposition segmentation in English and not for other tasks or languages. While evaluated positively on specific datasets, it may still exhibit errors and shares common LLM limitations regarding factual accuracy, common sense reasoning, and potential biases from its training data.