diag-scale diag_scale_1782792056

eval-logs/diag_scale_1782792056.txt

@@ -0,0 +1,13 @@
+RUN diag_scale_1782792056 convex+neural_feedback diag scaling curve
+==== POINT train_docs=5000 ====
+[N=5000] DIAG_TRAIN train_ppl=49.2870 train_docs=5000 vocab=1056 merges=800
+[N=5000]   expert[0]=freq     ppl=      500.51 acc=0.0102 mean_logit_spread=  10.444 gate_weight=0.0056
+[N=5000]   expert[1]=bigram   ppl=       84.67 acc=0.1022 mean_logit_spread=  16.011 gate_weight=3.1947
+[N=5000]   expert[2]=trigram  ppl=       54.06 acc=0.2495 mean_logit_spread=  14.866 gate_weight=4.9403
+[N=5000]   expert[3]=skip_bi  ppl=      278.23 acc=0.0438 mean_logit_spread=  14.374 gate_weight=0.0045
+[N=5000]   expert[4]=skip_tri ppl=      118.16 acc=0.1797 mean_logit_spread=  13.985 gate_weight=1.2362
+[N=5000]   expert[5]=neural_p ppl=      502.95 acc=0.0103 mean_logit_spread=   7.036 gate_weight=0.0054
+[N=5000]   expert[6]=neural_w ppl=      503.71 acc=0.0102 mean_logit_spread=   7.046 gate_weight=0.0054
+[N=5000] DIAG_MIXED mixed_ppl=41.50 mixed_acc=0.2672  (vs best standalone expert above)
+[N=5000] DIAG_NEURAL feature_dim=40 mean_across_ctx_std=0.01613 mean_feature_l2=1.6120 word_dec_l2=38.0250 phon_dec_l2=40.3025 embed_l2=10.5850
+[N=5000] point_elapsed_s=376 rc=0