diag-scale diag_scale_1782792056

eval-logs/diag_scale_1782792056.txt

@@ -23,3 +23,15 @@ RUN diag_scale_1782792056 convex+neural_feedback diag scaling curve
 [N=20000] DIAG_MIXED mixed_ppl=32.19 mixed_acc=0.2865  (vs best standalone expert above)
 [N=20000] DIAG_NEURAL feature_dim=40 mean_across_ctx_std=0.08081 mean_feature_l2=1.7973 word_dec_l2=54.2570 phon_dec_l2=58.7688 embed_l2=10.5850
 [N=20000] point_elapsed_s=1493 rc=0
+==== POINT train_docs=80000 ====
+[N=80000] DIAG_TRAIN train_ppl=30.0298 train_docs=80000 vocab=1056 merges=800
+[N=80000]   expert[0]=freq     ppl=      501.77 acc=0.0102 mean_logit_spread=  13.224 gate_weight=0.0188
+[N=80000]   expert[1]=bigram   ppl=       80.01 acc=0.1020 mean_logit_spread=  19.917 gate_weight=0.5350
+[N=80000]   expert[2]=trigram  ppl=       29.96 acc=0.2701 mean_logit_spread=  19.073 gate_weight=4.9433
+[N=80000]   expert[3]=skip_bi  ppl=      256.79 acc=0.0446 mean_logit_spread=  18.778 gate_weight=0.0183
+[N=80000]   expert[4]=skip_tri ppl=       48.97 acc=0.2176 mean_logit_spread=  18.071 gate_weight=2.4329
+[N=80000]   expert[5]=neural_p ppl=      449.32 acc=0.0148 mean_logit_spread=   8.546 gate_weight=0.0213
+[N=80000]   expert[6]=neural_w ppl=      466.46 acc=0.0131 mean_logit_spread=   8.161 gate_weight=0.0202
+[N=80000] DIAG_MIXED mixed_ppl=25.99 mixed_acc=0.3036  (vs best standalone expert above)
+[N=80000] DIAG_NEURAL feature_dim=40 mean_across_ctx_std=0.06980 mean_feature_l2=1.6484 word_dec_l2=97.6622 phon_dec_l2=103.8278 embed_l2=10.5850
+[N=80000] point_elapsed_s=6416 rc=0