21 pages, 1 figure, 9 tables. v2: retracted the B6 tool-id score of 0.08 (v1) after finding 3 benchmark harness bugs; under the fixed harness every B6/B7 metric is 0.0. Fixed a LoRA LR bug and a checkpoint-load-order bug. Added a 9-field gate + linear probe. Reframed the NoPE ablation as open, confound quantified. Code/checkpoints on HF.
翻译:暂无翻译