| Property | Value |
|---|---|
| Parameter Count | 11 Billion |
| Base Architecture | T5-XXL |
| Paper | arXiv:2303.13408 |
| Training Data | PAR3 Dataset |
What is dipper-paraphraser-xxl?
DIPPER (Discourse Paraphraser) is an advanced language model designed specifically for paraphrasing long-form text while maintaining contextual coherence. Built by fine-tuning T5-XXL, this 11B parameter model introduces innovative features for controlled text transformation, particularly useful for evading AI-generated text detectors while preserving original meaning.
Implementation Details
The model is implemented using the T5 architecture and trained on the PAR3 dataset, which contains multiple English translations of non-English novels. This unique training approach enables the model to understand and generate paragraph-level paraphrases while maintaining discourse-level coherence.
- Built on T5-XXL architecture with 11B parameters
- Trained on novel translations for robust paraphrasing capabilities
- Implements controllable diversity parameters
- Supports context-aware transformations
Core Capabilities
- Long-form text paraphrasing with context preservation
- Adjustable lexical diversity (0-100 scale)
- Controllable content reordering (0-100 scale)
- Paragraph-level transformation support
- Context-aware paraphrasing using input prompts
Frequently Asked Questions
Q: What makes this model unique?
DIPPER's uniqueness lies in its ability to paraphrase long-form text while maintaining discourse-level coherence, combined with precise control over both lexical diversity and content reordering. Unlike traditional paraphrasers that work at the sentence level, DIPPER operates on entire paragraphs while considering broader context.
Q: What are the recommended use cases?
The model is particularly suited for: content rephrasing while maintaining meaning, generating alternative versions of long-form text, creating variations of existing content with controlled diversity, and academic writing assistance. It's especially useful when working with paragraph-length texts that require maintaining contextual coherence.