RULER paper (arXiv:2404.06654v3) Yi-34B model analysis section and per-category effective-length tables

Section 6 (Model Analysis) and Appendix tables 13-16 of arXiv:2404.06654v3, which split the reported effective length BY TASK CATEGORY rather than reporting one number. For Yi-34B (200K claimed) the per-category effective lengths are: Retrieval (NIAH, 8 tasks) 64K; Multi-hop tracing (variable tracking) 16K; Aggregation (CWE/FWE) >128K; Question answering 128K. Same split for Llama-3.1-70B (128K claimed): 64K / 32K / 64K / 64K. For Llama-3.1-8B: 64K / 8K / 128K / 128K. Qwen2-72B: 64K / 32K / 64K / 64K. Command-R-plus: 64K / 32K / 64K / 128K. GLM-4-9B-1M: 64K / 8K / >128K / >128K. This is the load-bearing caveat for using a single 'effective length' number: the figure is category-dependent and the aggregate 13-task average hides a 4x spread within one model. Paper text on the Yi ablation: 'We evaluate Yi-34B-200K with increased input lengths (up to 256K) on more complex tasks' and 'training on longer sequences does not always lead to better performance on Ruler', plus 'larger model sizes positively correlate with better long-context capabilities'. Model list in Appendix A gives Yi's claimed length as 200K from 01-ai/Yi-34B-200K.