text stringlengths 941 4.74k | source stringclasses 1
value |
|---|---|
Preprint LONG WRITER : UNLEASHING 10,000+ W ORD GENERATION FROM LONG CONTEXT LLM S Yushi Bai1†, Jiajie Zhang1†, Xin Lv2, Linzhi Zheng1, Siqi Zhu1, Lei Hou1, Yuxiao Dong1, Jie Tang1, Juanzi Li1 1Tsinghua University2Zhipu AI ABSTRACT Current long context large language models (LLMs) can process inputs up to 100,000 token... | 2408.07055v1.pdf |
Preprint To address this limitation, we introduce Agent Write, a novel agent-based pipeline designed to lever-age off-the-shelf LLMs to automatically construct extended, coherent outputs (Sec. 3). Agent Write operates in two stages: First, it crafts a detailed writing plan outlining the structure and target word count ... | 2408.07055v1.pdf |
Preprint Figure 1: Long Writer-Ruler test demonstrates a maximum output length limitation of approxi-mately 2k words for all models tested. Figure 2: Long Writer-Ruler test of GLM-4-9B trained on SFT datasets of different maximum output lengths. Controlled experiment. We hypothesize that the common 2,000-word output le... | 2408.07055v1.pdf |
Preprint Write a 30000-word article on the history of the Roman Empire. Instruction LLMSTEP I: Plan Paragraph 1-Introduces the origins of the Roman Empire, including...-Word Count Requirement: 700 words Paragraph 2-Describe the founding of the Roman Empire, including...-Word Requirement: 800 words STEP II: Write... Par... | 2408.07055v1.pdf |
Preprint # Data in each subset Language Output type Chinese 60 Literature and Creative Writing 31 English 60 Academic and Monograph 22 Output length Popular Science 18 [0,500) 26 Functional Writing 17 [500,2000) 36 News Report 13 [2000,4000) 31 Community Forum 10 [4000,20000) 27 Education and Training 9 Average input l... | 2408.07055v1.pdf |
Preprint Overall [0, 500) [500, 2k) [2k, 4k) [4k, 20k) ¯S S l Sq Sl Sq Sl Sq Sl Sq Sl Sq GPT-4o 78. 6 65. 3 91. 8 91. 0 94. 6 91. 4 93. 6 65. 5 93. 0 5. 6 85. 3 +Agent Write 89. 1 86. 6 91. 6 91. 0 94. 6 91. 4 93. 6 77. 3 90. 2 86. 8 87. 5 +Parallel 88. 5 87. 2 88. 9 91. 0 94. 6 91. 4 93. 6 79. 2 85. 6 87. 3 80. 9 Tabl... | 2408.07055v1.pdf |
Preprint irrelevant identifiers like “paragraph 1”, “paragraph 2”, etc., that the model might have added at the beginning of each output section. We call our final obtained long output dataset “ longwriter-6k ”. In model training, to ensure the model's general capabilities, we combine longwriter-6k with gen-eral SFT da... | 2408.07055v1.pdf |
Preprint Overall [0, 500) [500, 2k) [2k, 4k) [4k, 20k) ¯S S l Sq Sl Sq Sl Sq Sl Sq Sl Sq Proprietary models Claude 3. 5 Sonnet 80. 7 73. 7 87. 7 87. 0 92. 5 93. 6 90. 4 81. 3 86. 6 26. 0 80. 9 GPT-4 Turbo 67. 3 47. 9 86. 6 92. 0 90. 2 81. 2 90. 7 12. 3 85. 5 0 78. 7 GPT-4o mini 77. 6 64. 9 90. 3 92. 8 95. 4 91. 7 93. 1... | 2408.07055v1.pdf |
Preprint Figure 7: Cumulative average NLL loss of GLM-4-9B and Llama-3. 1-8B at different positions of Long Writer models' outputs. Figure 8: Long Write-Ruler test results of Long-Writer models, showing their maximum genera-tion lengths between 10k-20k words. the requested length. The latter conclusion has also been re... | 2408.07055v1.pdf |
Preprint Overall [0, 500) [500, 2k) [2k, 4k) [4k, 20k) ¯S S l Sq Sl Sq Sl Sq Sl Sq Sl Sq Long Writer-9B 80. 5 78. 6 82. 3 83. 9 86. 2 75. 6 84. 8 76. 0 80. 2 80. 3 77. 3-Long Writer-6k data 62. 6 48. 1 77. 1 83. 8 85. 1 77. 8 79. 6 25. 7 71. 9 0 71. 9 w/ Plan-augmented data 81. 4 80. 9 81. 8 85. 9 84. 0 79. 4 82. 3 78.... | 2408.07055v1.pdf |
Preprint content”. This includes zero-shot extension methods (Han et al., 2023; Xiao et al., 2023; Zhang et al., 2024a; Jin et al., 2024; An et al., 2024), as well as methods that involve fine-tuning the model on longer sequences to achieve a longer memory (Chen et al., 2023a; Peng et al., 2023; Xiong et al., 2024; Che... | 2408.07055v1.pdf |
Preprint Collin Burns, Pavel Izmailov, Jan Hendrik Kirchner, Bowen Baker, Leo Gao, Leopold Aschenbren-ner, Yining Chen, Adrien Ecoffet, Manas Joglekar, Jan Leike, et al. Weak-to-strong general-ization: Eliciting strong capabilities with weak supervision. ar Xiv preprint ar Xiv:2312. 09390, 2023. Tianle Cai, Yuhong Li, ... | 2408.07055v1.pdf |
Preprint Zhenyu Hou, Yiin Niu, Zhengxiao Du, Xiaohan Zhang, Xiao Liu, Aohan Zeng, Qinkai Zheng, Minlie Huang, Hongning Wang, Jie Tang, et al. Chatglm-rlhf: Practices of aligning large language models with human feedback. ar Xiv preprint ar Xiv:2404. 00934, 2024. Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zh... | 2408.07055v1.pdf |
Preprint Haoran Sun, Lixin Liu, Junjie Li, Fengyu Wang, Baohua Dong, Ran Lin, and Ruohui Huang. Conifer: Improving complex constrained instruction-following ability of large language models. ar Xiv preprint ar Xiv:2404. 02823, 2024. Lewis Tunstall, Edward Beeching, Nathan Lambert, Nazneen Rajani, Kashif Rasul, Younes B... | 2408.07055v1.pdf |
Preprint A M ODEL CARDS We list the details of our evaluated models in Table 5. Model name Model version Context window Max output tokens Claude 3. 5 Sonnet (Anthropic, 2024) claude-3-5-sonnet-20240620 200,000 tokens 4,096 tokens GPT-4 Turbo (Achiam et al., 2023) gpt-4-turbo-2024-04-09 128,000 tokens 4,096 tokens GPT-4... | 2408.07055v1.pdf |
Preprint ⟨Response ⟩ {Model response } ⟨/Response ⟩ Please evaluate the quality of the response. You must first provide a brief analysis of its quality, then give a comprehensive analysis with scores for each dimension. The output must strictly follow the JSON format: {“Analysis”:..., “Relevance”:..., “Accuracy”:..., “... | 2408.07055v1.pdf |
Preprint [0, 500) [500, 2k) [2k, 4k) [4k, 20k) Mean Median Mean Median Mean Median Mean Median Required Length 294 300 894 800 2,477 2,400 8,000 6,000 Proprietary models Claude 3. 5 Sonnet 357 342 927 877 1,891 1,896 2,399 2,881 GPT-4 Turbo 291 294 660 626 778 785 907 701 GPT-4o mini 331 317 884 848 2,218 1,455 1,631 1... | 2408.07055v1.pdf |
README.md exists but content is empty.
- Downloads last month
- 17