{ "normal": { "all": { "targets": 169144, "max_abs_delta": 0.0, "nll": 4.144884785078792, "delta_nats": 0.0, "ppl_ratio": 1.0, "doc_bootstrap_delta_95pct": [ 0.0, 0.0 ] }, "length_10-16": { "targets": 56601, "max_abs_delta": 0.0, "nll": 3.639815218349103, "delta_nats": 0.0, "ppl_ratio": 1.0, "doc_bootstrap_delta_95pct": [ 0.0, 0.0 ] }, "before_first": { "targets": 83341, "max_abs_delta": 0.0, "nll": 3.6582595163808413, "delta_nats": 0.0, "ppl_ratio": 1.0, "doc_bootstrap_delta_95pct": [ 0.0, 0.0 ] }, "after_first": { "targets": 85803, "max_abs_delta": 0.0, "nll": 4.6175470057303505, "delta_nats": 0.0, "ppl_ratio": 1.0, "doc_bootstrap_delta_95pct": [ 0.0, 0.0 ] }, "after_second": { "targets": 15328, "max_abs_delta": 0.0, "nll": 4.777278997806058, "delta_nats": 0.0, "ppl_ratio": 1.0, "doc_bootstrap_delta_95pct": [ 0.0, 0.0 ] }, "length_17-64": { "targets": 48171, "max_abs_delta": 0.0, "nll": 4.587060177184653, "delta_nats": 0.0, "ppl_ratio": 1.0, "doc_bootstrap_delta_95pct": [ 0.0, 0.0 ] }, "length_2-9": { "targets": 25421, "max_abs_delta": 0.0, "nll": 3.312832962033229, "delta_nats": 0.0, "ppl_ratio": 1.0, "doc_bootstrap_delta_95pct": [ 0.0, 0.0 ] }, "length_65-128": { "targets": 22018, "max_abs_delta": 0.0, "nll": 4.893068085736666, "delta_nats": 0.0, "ppl_ratio": 1.0, "doc_bootstrap_delta_95pct": [ 0.0, 0.0 ] }, "length_129+": { "targets": 16933, "max_abs_delta": 0.0, "nll": 4.851522782663743, "delta_nats": 0.0, "ppl_ratio": 1.0, "doc_bootstrap_delta_95pct": [ 0.0, 0.0 ] } }, "C_read_off": { "all": { "targets": 169144, "max_abs_delta": 9.837771174497902, "nll": 4.145392488348739, "delta_nats": 0.0005077032699674228, "ppl_ratio": 1.0005078321730865, "doc_bootstrap_delta_95pct": [ -0.00025010107714061553, 0.0014034107223756765 ] }, "length_10-16": { "targets": 56601, "max_abs_delta": 4.806448265910149, "nll": 3.638001250299956, "delta_nats": -0.0018139680491534834, "ppl_ratio": 0.9981876761965349, "doc_bootstrap_delta_95pct": [ -0.00262191558572389, -0.0008871909215438053 ] }, "before_first": { "targets": 83341, "max_abs_delta": 0.0, "nll": 3.6582595163808413, "delta_nats": 0.0, "ppl_ratio": 1.0, "doc_bootstrap_delta_95pct": [ 0.0, 0.0 ] }, "after_first": { "targets": 85803, "max_abs_delta": 9.837771174497902, "nll": 4.618547844417748, "delta_nats": 0.001000838687404517, "ppl_ratio": 1.0010013396935717, "doc_bootstrap_delta_95pct": [ -0.0006520181965873735, 0.0027062239328431454 ] }, "after_second": { "targets": 15328, "max_abs_delta": 0.9968395233154297, "nll": 4.779742083394126, "delta_nats": 0.002463085588067861, "ppl_ratio": 1.0024661214754134, "doc_bootstrap_delta_95pct": [ 0.0007099402930980536, 0.004136419760806341 ] }, "length_17-64": { "targets": 48171, "max_abs_delta": 9.837771174497902, "nll": 4.588660769801097, "delta_nats": 0.00160059261644365, "ppl_ratio": 1.0016018742485047, "doc_bootstrap_delta_95pct": [ -0.0007734870164671116, 0.004409285206532606 ] }, "length_2-9": { "targets": 25421, "max_abs_delta": 0.0, "nll": 3.312832962033229, "delta_nats": 0.0, "ppl_ratio": 1.0, "doc_bootstrap_delta_95pct": [ 0.0, 0.0 ] }, "length_65-128": { "targets": 22018, "max_abs_delta": 1.317411184310913, "nll": 4.896818332041356, "delta_nats": 0.003750246304688991, "ppl_ratio": 1.0037572872774045, "doc_bootstrap_delta_95pct": [ 0.0024134010107044013, 0.005072407909455955 ] }, "length_129+": { "targets": 16933, "max_abs_delta": 0.9968395233154297, "nll": 4.853227873160534, "delta_nats": 0.0017050904967906946, "ppl_ratio": 1.0017065449901552, "doc_bootstrap_delta_95pct": [ 5.13874762613377e-05, 0.003163192791394106 ] } }, "D_read_off": { "all": { "targets": 169144, "max_abs_delta": 14.808307343628258, "nll": 4.563928485893732, "delta_nats": 0.4190437008148539, "ppl_ratio": 1.5205068007244975, "doc_bootstrap_delta_95pct": [ 0.4084816376528276, 0.42974781810001217 ] }, "length_10-16": { "targets": 56601, "max_abs_delta": 14.808307343628258, "nll": 3.947487452312927, "delta_nats": 0.3076722339636419, "ppl_ratio": 1.3602550704497973, "doc_bootstrap_delta_95pct": [ 0.29860700436242454, 0.3170346994210013 ] }, "before_first": { "targets": 83341, "max_abs_delta": 9.79684591293335, "nll": 3.8673202743180943, "delta_nats": 0.2090607579371381, "ppl_ratio": 1.2325198816208498, "doc_bootstrap_delta_95pct": [ 0.20296126497060282, 0.2153221963653742 ] }, "after_first": { "targets": 85803, "max_abs_delta": 14.808307343628258, "nll": 5.240548475415436, "delta_nats": 0.6230014696850796, "ppl_ratio": 1.864515939768767, "doc_bootstrap_delta_95pct": [ 0.6098227767522754, 0.6367022716297845 ] }, "after_second": { "targets": 15328, "max_abs_delta": 9.862558734603226, "nll": 5.4398346708231395, "delta_nats": 0.6625556730170808, "ppl_ratio": 1.939743354864635, "doc_bootstrap_delta_95pct": [ 0.6337338080399308, 0.6941607536956912 ] }, "length_17-64": { "targets": 48171, "max_abs_delta": 13.812182898342144, "nll": 5.070832321187806, "delta_nats": 0.48377214400309654, "ppl_ratio": 1.6221819796464214, "doc_bootstrap_delta_95pct": [ 0.4675586082944131, 0.49992120057822614 ] }, "length_2-9": { "targets": 25421, "max_abs_delta": 9.79684591293335, "nll": 3.5187845186227595, "delta_nats": 0.20595155658954448, "ppl_ratio": 1.228693680441276, "doc_bootstrap_delta_95pct": [ 0.1965656321666891, 0.2159041335973384 ] }, "length_65-128": { "targets": 22018, "max_abs_delta": 10.74917882680893, "nll": 5.536569118899567, "delta_nats": 0.6435010331629004, "ppl_ratio": 1.9031321581362317, "doc_bootstrap_delta_95pct": [ 0.6137474662493506, 0.6740910758874403 ] }, "length_129+": { "targets": 16933, "max_abs_delta": 9.862558734603226, "nll": 5.486748873671533, "delta_nats": 0.6352260910077897, "ppl_ratio": 1.8874488284296478, "doc_bootstrap_delta_95pct": [ 0.6040921677702079, 0.6676251294704127 ] } }, "interpretation": "within-model read interventions, not retrained ablation, not six-axis ranking; corpus sampled before this model, 100111-word dev subset; document bootstrap ignores cross-document duplicate/dependence" }