Antislop: A Framework for Eliminating Repetitive Patterns in Language Models
6 months ago
- #AI Ethics
- #Machine Learning
- #Natural Language Processing
- Introduction of 'slop' as repetitive phraseology in LLM outputs degrading quality and recognizability.
- Presentation of Antislop framework with three key innovations: Antislop Sampler, automated slop profiling pipeline, and FTPO fine-tuning method.
- Demonstration of slop patterns appearing over 1,000x more frequently in LLM outputs than human text.
- Effectiveness of Antislop Sampler in suppressing over 8,000 patterns without quality loss.
- Superior performance of FTPO with 90% slop reduction and maintained or improved cross-domain evaluation results.
- Comparison showing DPO's degradation in writing quality and lexical diversity versus FTPO.
- Release of all code and results under MIT license for public access.