AI Autocorrect Engine
| Property | Value |
|---|---|
| Author | DavidAU |
| License | Open License (SillyTavern) |
| Framework | SillyTavern Plugin |
| Application Type | Text Generation Enhancement |
What is AI_Autocorrect?
This is a sophisticated AI enhancement engine designed to improve text generation quality across all GGUF models, particularly focusing on real-time correction and creative enhancement. Built with over 30 years of programming experience and extensive model testing, it operates as a software patch within the SillyTavern framework.
Implementation Details
The system implements two primary mechanisms: Auto-Correction and Reconsider systems. It samples the generation stream 50 times per second, actively correcting issues like repetitions, incoherence, and token errors. The engine uses dynamic parameter adjustment, including temperature and top-k sampling modifications, to optimize output quality.
- Real-time correction of letter, word, sentence, and paragraph repetitions
- Automatic handling of model instabilities and generation errors
- Support for low-quantization model optimization (IQ1s, IQ2s, q2k)
- Five-level adjustment system for temperature and top-k parameters
Core Capabilities
- Active error detection and correction without user intervention
- Paragraph and sentence level reconsideration system
- Dynamic parameter adjustment for enhanced creativity
- Support for multiple model types and quantization levels
- Integration with various AI/LLM applications including LMStudio, Koboldcpp, and Text Generation WebUI
Frequently Asked Questions
Q: What makes this model unique?
This engine stands out for its ability to actively modify and fine-tune generation in real-time, creating a two-way partnership between the engine and AI/LLM. It allows models to operate at full power without restrictive sampling parameters while maintaining coherent output.
Q: What are the recommended use cases?
The engine is particularly useful for enhancing the performance of creative/unstable models, low-quantization models, and any scenario requiring high-quality text generation with minimal manual intervention. It's especially valuable for class 2-4 models that typically require careful parameter management.
