From 0d167497b75031fae67b71fa1e7ab12e57c761bb Mon Sep 17 00:00:00 2001 From: YurenHao0426 Date: Mon, 10 Aug 2026 10:50:16 -0500 Subject: results: close BabyAI population predictor rescue --- results/babyai_shared/population_p3_analysis.json | 176 ++++++++++++++++++++++ 1 file changed, 176 insertions(+) create mode 100644 results/babyai_shared/population_p3_analysis.json (limited to 'results/babyai_shared/population_p3_analysis.json') diff --git a/results/babyai_shared/population_p3_analysis.json b/results/babyai_shared/population_p3_analysis.json new file mode 100644 index 0000000..08a9c7c --- /dev/null +++ b/results/babyai_shared/population_p3_analysis.json @@ -0,0 +1,176 @@ +{ + "checks": { + "all_seed_gates_pass": false, + "r2_at_least_0p8_every_seed_epoch_layer": false, + "residual_ratio_at_most_0p25_every_seed_epoch_layer": false, + "zero_action_or_teaching_observations": true + }, + "decision": "Close the population-linear predictor rescue. It is highly predictive at initialization but fails the frozen early-layer threshold after clean task learning; do not run another PickupLoc downstream endpoint.", + "epoch_layer_means_across_seeds": [ + { + "epoch": 0, + "layer": 0, + "mean_holdout_r2": 0.972669243812561, + "mean_holdout_residual_ratio": 0.10559078057607015 + }, + { + "epoch": 0, + "layer": 1, + "mean_holdout_r2": 0.9938111305236816, + "mean_holdout_residual_ratio": 0.05075520773728689 + }, + { + "epoch": 0, + "layer": 2, + "mean_holdout_r2": 0.9965593020121256, + "mean_holdout_residual_ratio": 0.037968567262093224 + }, + { + "epoch": 0, + "layer": 3, + "mean_holdout_r2": 0.9977092742919922, + "mean_holdout_residual_ratio": 0.03117166894177596 + }, + { + "epoch": 1, + "layer": 0, + "mean_holdout_r2": 0.8532736500104269, + "mean_holdout_residual_ratio": 0.24617203076680502 + }, + { + "epoch": 1, + "layer": 1, + "mean_holdout_r2": 0.9774375756581625, + "mean_holdout_residual_ratio": 0.09696057438850403 + }, + { + "epoch": 1, + "layer": 2, + "mean_holdout_r2": 0.9829351902008057, + "mean_holdout_residual_ratio": 0.08451603849728902 + }, + { + "epoch": 1, + "layer": 3, + "mean_holdout_r2": 0.9884803891181946, + "mean_holdout_residual_ratio": 0.07001137733459473 + }, + { + "epoch": 5, + "layer": 0, + "mean_holdout_r2": 0.7852290074030558, + "mean_holdout_residual_ratio": 0.2982979516188304 + }, + { + "epoch": 5, + "layer": 1, + "mean_holdout_r2": 0.9404525359471639, + "mean_holdout_residual_ratio": 0.15772825479507446 + }, + { + "epoch": 5, + "layer": 2, + "mean_holdout_r2": 0.9551665385564169, + "mean_holdout_residual_ratio": 0.13717531661192575 + }, + { + "epoch": 5, + "layer": 3, + "mean_holdout_r2": 0.9778080781300863, + "mean_holdout_residual_ratio": 0.09727375209331512 + }, + { + "epoch": 10, + "layer": 0, + "mean_holdout_r2": 0.7818694313367208, + "mean_holdout_residual_ratio": 0.30070162812868756 + }, + { + "epoch": 10, + "layer": 1, + "mean_holdout_r2": 0.9168808062871298, + "mean_holdout_residual_ratio": 0.1869215965270996 + }, + { + "epoch": 10, + "layer": 2, + "mean_holdout_r2": 0.9389026165008545, + "mean_holdout_residual_ratio": 0.16029799977938333 + }, + { + "epoch": 10, + "layer": 3, + "mean_holdout_r2": 0.9738152623176575, + "mean_holdout_residual_ratio": 0.10578999916712443 + }, + { + "epoch": 20, + "layer": 0, + "mean_holdout_r2": 0.8234108686447144, + "mean_holdout_residual_ratio": 0.2713727255662282 + }, + { + "epoch": 20, + "layer": 1, + "mean_holdout_r2": 0.9010587930679321, + "mean_holdout_residual_ratio": 0.20385040839513144 + }, + { + "epoch": 20, + "layer": 2, + "mean_holdout_r2": 0.9290995995203654, + "mean_holdout_residual_ratio": 0.17255698641141257 + }, + { + "epoch": 20, + "layer": 3, + "mean_holdout_r2": 0.9702359239260355, + "mean_holdout_residual_ratio": 0.11271155128876369 + }, + { + "epoch": 40, + "layer": 0, + "mean_holdout_r2": 0.7636886239051819, + "mean_holdout_residual_ratio": 0.31424906849861145 + }, + { + "epoch": 40, + "layer": 1, + "mean_holdout_r2": 0.8482151031494141, + "mean_holdout_residual_ratio": 0.2532327820857366 + }, + { + "epoch": 40, + "layer": 2, + "mean_holdout_r2": 0.8937068382898966, + "mean_holdout_residual_ratio": 0.21221035222212473 + }, + { + "epoch": 40, + "layer": 3, + "mean_holdout_r2": 0.9683659275372823, + "mean_holdout_residual_ratio": 0.11610875278711319 + } + ], + "gate": "fail", + "maximum_holdout_residual_context_rms_ratio": { + "context_rms": 0.20538413524627686, + "epoch": 40, + "layer": 0, + "mean_per_cell_r2": 0.7517150640487671, + "observations": 4096, + "residual_context_rms_ratio": 0.32660192251205444, + "seed": 4103 + }, + "minimum_holdout_r2": { + "context_rms": 0.20538413524627686, + "epoch": 40, + "layer": 0, + "mean_per_cell_r2": 0.7517150640487671, + "observations": 4096, + "residual_context_rms_ratio": 0.32660192251205444, + "seed": 4103 + }, + "raw_sdil_rollout_or_test_outcomes_read": false, + "stage": "babyai_population_p3_analysis" +} -- cgit v1.2.3