Summary of Longwriter: Unleashing 10,000+ Word Generation From Long Context Llms, by Yushi Bai et al.
LongWriter: Unleashing 10,000+ Word Generation from Long Context LLMs
by Yushi Bai, Jiajie Zhang, Xin Lv, Linzhi Zheng, Siqi Zhu, Lei Hou, Yuxiao Dong, Jie Tang, Juanzi Li
First submitted to arxiv on: 13 Aug 2024
Categories
- Main: Computation and Language (cs.CL)
- Secondary: Machine Learning (cs.LG)
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary This paper tackles the limitation of current long context large language models (LLMs) in generating outputs exceeding a modest length. Through experiments, researchers find that the model’s effective generation length is bounded by the sample it has seen during supervised fine-tuning (SFT). The scarcity of long-output examples in existing SFT datasets is the primary cause. To overcome this limitation, the authors introduce AgentWrite, an agent-based pipeline decomposing ultra-long generation tasks into subtasks. This enables off-the-shelf LLMs to generate coherent outputs exceeding 20,000 words. By leveraging the AgentWrite pipeline and constructing a dataset (LongWriter-6k) containing 6,000 SFT data with output lengths ranging from 2k to 32k words, researchers successfully scale the output length of existing models to over 10,000 words while maintaining quality. The authors also develop LongBench-Write, a benchmark for evaluating ultra-long generation capabilities. Their 9B parameter model, improved through DPO, achieves state-of-the-art performance on this benchmark, outperforming even larger proprietary models. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This paper helps computers understand and write long texts better. Right now, these machines can only write short texts, but scientists want to make them able to write longer ones. To do this, they created a new way of training the computers (called AgentWrite) that breaks down big tasks into smaller ones. This lets them write really long texts that are still understandable. The scientists also made a special dataset with lots of examples of writing long texts, which helps the computers learn how to do it better. With this new method and dataset, they were able to make the computers write even longer texts while keeping them good quality. |
Keywords
* Artificial intelligence * Fine tuning * Supervised
