Self-Lengthen - An iterative training framework launched by Alibaba's Qianwen to improve output length
Self-Lengthen is an innovative iterative training framework developed by Alibaba's Qianwen team, which enhances the ability of large language models (LLMs) to generate long texts. The framework is based on two roles: a generator and an expander, working together to generate...
What is Self-Lengthen?
Self-Lengthen is an innovative iterative training framework developed by Alibaba's Qianwen team, which enhances the ability of large language models (LLMs) to generate long texts. The framework is based on two roles: a generator and an expander. The generator produces the initial response, while the expander splits and expands the response to produce longer text. The entire process iterates continuously, gradually training the model to handle longer outputs. Self-Lengthen requires no additional data or proprietary models, effectively addressing the training deficiencies in long text generation by leveraging the inherent knowledge and skills of LLMs.
Main functions of Self-Lengthen
- Increase output lengthThis allows LLMs to generate longer text outputs than traditional training methods.
- Maintain content qualityWhile extending the text length, it maintains or even enhances the coherence and relevance of the generated content.
- No additional data requiredIt does not rely on external data sources or proprietary models, but is based on the knowledge and skills inherent in the model.
- Iterative trainingThe ability of the model to process long texts is gradually improved through an iterative process.
- flexibilityIt can be applied to a variety of long text generation tasks, including literary creation and academic research.
Self-Lengthen's technical principle
- Generator and Extender:
- generator: Responsible for generating the initial short text response.
- Expander: Take the generator's output as input and expand it into a long text.
- Iterative training process:
- Through repeated iterations, the generator and expander's ability to process long texts is gradually increased.
- In each iteration, the expander attempts to extend the generator's output to be longer, fine-tunes the generator with the longer output, and directly generates longer text.
- Instruction Augmentation: Expand and diversify training instructions using self-guided techniques to better guide the model in generating long texts.
- Two-phase expansion method:
- Phase 1The expander expands the first half of the generator's output.
- Phase TwoThe results of the first stage of expansion guide the expansion of the remaining parts, thus expanding the entire text.
- Fine-tuning model: Use an extension-based generator and extender to generate longer text, making it easier to generate longer text in future iterations.
- Quality controlThe system uses rules and evaluation mechanisms to ensure the quality of generated long texts and avoids repetitive and meaningless expansions.
Self-Lengthen project address
- GitHub repository:https://github.com/QwenLM/Self-Lengthen
- arXiv technical paper:https://arxiv.org/pdf/2410.23933
Application scenarios of Self-Lengthen
- Creative WritingUsed to generate long literary works such as novels, stories, and scripts.
- academic researchIt assists scholars and researchers in writing academic papers, technical reports, and research proposals.
- News mediaUsed in writing news reports, in-depth articles, and feature reports to provide comprehensive content coverage.
- Educational content developmentCreate educational materials, course content, and textbooks, and provide in-depth teaching resources.
- Business copywritingWriting marketing copy, advertising content, and business plans, among other business documents.