refactoring

31 files changed, 493 insertions, 757 deletions

diff --git a/training/dtrain/examples/parallelized/work/out.0.0 b/training/dtrain/examples/parallelized/work/out.0.0
index 9154c906..77749404 100644
--- a/training/dtrain/examples/parallelized/work/out.0.0
+++ b/training/dtrain/examples/parallelized/work/out.0.0
@@ -1,65 +1,43 @@
-                cdec cfg 'cdec.ini'
 Loading the LM will be faster if you build a binary file.
 Reading ../standard/nc-wmt11.en.srilm.gz
 ----5---10---15---20---25---30---35---40---45---50---55---60---65---70---75---80---85---90---95--100
 ****************************************************************************************************
-Seeding random number sequence to 4087834873
-
 dtrain
 Parameters:
                        k 100
                        N 4
                        T 1
-                   batch 0
-                  scorer 'stupid_bleu'
-             sample from 'kbest'
-                  filter 'uniq'
            learning rate 0.0001
-                   gamma 0
-             loss margin 1
-       faster perceptron 0
-                   pairs 'XYX'
-                   hi lo 0.1
-          pair threshold 0
-          select weights 'last'
-                  l1 reg 0 'none'
-                    pclr no
-               max pairs 4294967295
-                  repeat 1
-                cdec cfg 'cdec.ini'
-                   input ''
+            error margin 1
+                  l1 reg 0
+            decoder conf 'cdec.ini'
+                   input 'work/shard.0.0.in'
                   output 'work/weights.0.0'
-(a dot represents 10 inputs)
+(a dot per input)
 Iteration #1 of 1.
-  3
+ .... 3
 WEIGHTS
-              Glue = +0.257
-       WordPenalty = +0.026926
-     LanguageModel = +0.67342
- LanguageModel_OOV = -0.046
-     PhraseModel_0 = +0.25329
-     PhraseModel_1 = +0.20036
-     PhraseModel_2 = +0.00060731
-     PhraseModel_3 = +0.65578
-     PhraseModel_4 = +0.47916
-     PhraseModel_5 = +0.004
-     PhraseModel_6 = +0.1829
-       PassThrough = -0.082
+              Glue = +0.3404
+       WordPenalty = -0.017632
+     LanguageModel = +0.72958
+ LanguageModel_OOV = -0.235
+     PhraseModel_0 = -0.43721
+     PhraseModel_1 = +1.01
+     PhraseModel_2 = +1.3525
+     PhraseModel_3 = -0.25541
+     PhraseModel_4 = -0.78115
+     PhraseModel_5 = +0
+     PhraseModel_6 = -0.3681
+       PassThrough = -0.3304
         ---
-       1best avg score: 0.04518 (+0.04518)
- 1best avg model score: 32.803 (+32.803)
-           avg # pairs: 1266.3
-        avg # rank err: 857
-     avg # margin viol: 386.67
-       k-best loss imp: 100%
-    non0 feature count: 12
+       1best avg score: 0.19474 (+0.19474)
+ 1best avg model score: 0.52232
+           avg # pairs: 2513
+   non-0 feature count: 11
            avg list sz: 100
-           avg f count: 10.853
-(time 0.47 min, 9.3 s/S)
-
-Writing weights file to 'work/weights.0.0' ...
-done
+           avg f count: 11.42
+(time 0.32 min, 6 s/S)
 
 ---
-Best iteration: 1 [SCORE 'stupid_bleu'=0.04518].
-This took 0.46667 min.
+Best iteration: 1 [GOLD = 0.19474].
+This took 0.31667 min.
diff --git a/training/dtrain/examples/parallelized/work/out.0.1 b/training/dtrain/examples/parallelized/work/out.0.1
index 0dbc7bd3..d0dee623 100644
--- a/training/dtrain/examples/parallelized/work/out.0.1
+++ b/training/dtrain/examples/parallelized/work/out.0.1
@@ -1,66 +1,44 @@
-                cdec cfg 'cdec.ini'
 Loading the LM will be faster if you build a binary file.
 Reading ../standard/nc-wmt11.en.srilm.gz
 ----5---10---15---20---25---30---35---40---45---50---55---60---65---70---75---80---85---90---95--100
 ****************************************************************************************************
-Seeding random number sequence to 2283043509
-
 dtrain
 Parameters:
                        k 100
                        N 4
                        T 1
-                   batch 0
-                  scorer 'stupid_bleu'
-             sample from 'kbest'
-                  filter 'uniq'
            learning rate 0.0001
-                   gamma 0
-             loss margin 1
-       faster perceptron 0
-                   pairs 'XYX'
-                   hi lo 0.1
-          pair threshold 0
-          select weights 'last'
-                  l1 reg 0 'none'
-                    pclr no
-               max pairs 4294967295
-                  repeat 1
-                cdec cfg 'cdec.ini'
-                   input ''
+            error margin 1
+                  l1 reg 0
+            decoder conf 'cdec.ini'
+                   input 'work/shard.0.0.in'
                   output 'work/weights.0.1'
               weights in 'work/weights.0'
-(a dot represents 10 inputs)
+(a dot per input)
 Iteration #1 of 1.
-  3
+ .... 3
 WEIGHTS
-              Glue = -0.17905
-       WordPenalty = +0.062126
-     LanguageModel = +0.66825
- LanguageModel_OOV = -0.15248
-     PhraseModel_0 = -0.55811
-     PhraseModel_1 = +0.12741
-     PhraseModel_2 = +0.60388
-     PhraseModel_3 = -0.44464
-     PhraseModel_4 = -0.63137
-     PhraseModel_5 = -0.0084
-     PhraseModel_6 = -0.20165
-       PassThrough = -0.23468
+              Glue = -0.40908
+       WordPenalty = +0.12967
+     LanguageModel = +0.39892
+ LanguageModel_OOV = -0.6314
+     PhraseModel_0 = -0.63992
+     PhraseModel_1 = +0.74198
+     PhraseModel_2 = +1.3096
+     PhraseModel_3 = -0.1216
+     PhraseModel_4 = -1.2274
+     PhraseModel_5 = +0.02435
+     PhraseModel_6 = -0.21093
+       PassThrough = -0.66155
         ---
-       1best avg score: 0.14066 (+0.14066)
- 1best avg model score: -37.614 (-37.614)
-           avg # pairs: 1244.7
-        avg # rank err: 728
-     avg # margin viol: 516.67
-       k-best loss imp: 100%
-    non0 feature count: 12
+       1best avg score: 0.15735 (+0.15735)
+ 1best avg model score: 46.831
+           avg # pairs: 2132.3
+   non-0 feature count: 12
            avg list sz: 100
-           avg f count: 11.507
-(time 0.45 min, 9 s/S)
-
-Writing weights file to 'work/weights.0.1' ...
-done
+           avg f count: 10.64
+(time 0.38 min, 7 s/S)
 
 ---
-Best iteration: 1 [SCORE 'stupid_bleu'=0.14066].
-This took 0.45 min.
+Best iteration: 1 [GOLD = 0.15735].
+This took 0.38333 min.
diff --git a/training/dtrain/examples/parallelized/work/out.0.2 b/training/dtrain/examples/parallelized/work/out.0.2
index fcecc7e1..9c4b110b 100644
--- a/training/dtrain/examples/parallelized/work/out.0.2
+++ b/training/dtrain/examples/parallelized/work/out.0.2
@@ -1,66 +1,44 @@
-                cdec cfg 'cdec.ini'
 Loading the LM will be faster if you build a binary file.
 Reading ../standard/nc-wmt11.en.srilm.gz
 ----5---10---15---20---25---30---35---40---45---50---55---60---65---70---75---80---85---90---95--100
 ****************************************************************************************************
-Seeding random number sequence to 3693132895
-
 dtrain
 Parameters:
                        k 100
                        N 4
                        T 1
-                   batch 0
-                  scorer 'stupid_bleu'
-             sample from 'kbest'
-                  filter 'uniq'
            learning rate 0.0001
-                   gamma 0
-             loss margin 1
-       faster perceptron 0
-                   pairs 'XYX'
-                   hi lo 0.1
-          pair threshold 0
-          select weights 'last'
-                  l1 reg 0 'none'
-                    pclr no
-               max pairs 4294967295
-                  repeat 1
-                cdec cfg 'cdec.ini'
-                   input ''
+            error margin 1
+                  l1 reg 0
+            decoder conf 'cdec.ini'
+                   input 'work/shard.0.0.in'
                   output 'work/weights.0.2'
               weights in 'work/weights.1'
-(a dot represents 10 inputs)
+(a dot per input)
 Iteration #1 of 1.
-  3
+ .... 3
 WEIGHTS
-              Glue = -0.019275
-       WordPenalty = +0.022192
-     LanguageModel = +0.40688
- LanguageModel_OOV = -0.36397
-     PhraseModel_0 = -0.36273
-     PhraseModel_1 = +0.56432
-     PhraseModel_2 = +0.85638
-     PhraseModel_3 = -0.20222
-     PhraseModel_4 = -0.48295
-     PhraseModel_5 = +0.03145
-     PhraseModel_6 = -0.26092
-       PassThrough = -0.38122
+              Glue = -0.44422
+       WordPenalty = +0.1032
+     LanguageModel = +0.66474
+ LanguageModel_OOV = -0.62252
+     PhraseModel_0 = -0.59993
+     PhraseModel_1 = +0.78992
+     PhraseModel_2 = +1.3149
+     PhraseModel_3 = +0.21434
+     PhraseModel_4 = -1.0174
+     PhraseModel_5 = +0.02435
+     PhraseModel_6 = -0.18452
+       PassThrough = -0.65268
         ---
-       1best avg score: 0.18982 (+0.18982)
- 1best avg model score: 1.7096 (+1.7096)
-           avg # pairs: 1524.3
-        avg # rank err: 813.33
-     avg # margin viol: 702.67
-       k-best loss imp: 100%
-    non0 feature count: 12
+       1best avg score: 0.24722 (+0.24722)
+ 1best avg model score: 61.971
+           avg # pairs: 2017.7
+   non-0 feature count: 12
            avg list sz: 100
-           avg f count: 11.32
-(time 0.53 min, 11 s/S)
-
-Writing weights file to 'work/weights.0.2' ...
-done
+           avg f count: 10.42
+(time 0.3 min, 6 s/S)
 
 ---
-Best iteration: 1 [SCORE 'stupid_bleu'=0.18982].
-This took 0.53333 min.
+Best iteration: 1 [GOLD = 0.24722].
+This took 0.3 min.
diff --git a/training/dtrain/examples/parallelized/work/out.1.0 b/training/dtrain/examples/parallelized/work/out.1.0
index 595dfc94..3dc4dca6 100644
--- a/training/dtrain/examples/parallelized/work/out.1.0
+++ b/training/dtrain/examples/parallelized/work/out.1.0
@@ -1,65 +1,43 @@
-                cdec cfg 'cdec.ini'
 Loading the LM will be faster if you build a binary file.
 Reading ../standard/nc-wmt11.en.srilm.gz
 ----5---10---15---20---25---30---35---40---45---50---55---60---65---70---75---80---85---90---95--100
 ****************************************************************************************************
-Seeding random number sequence to 859043351
-
 dtrain
 Parameters:
                        k 100
                        N 4
                        T 1
-                   batch 0
-                  scorer 'stupid_bleu'
-             sample from 'kbest'
-                  filter 'uniq'
            learning rate 0.0001
-                   gamma 0
-             loss margin 1
-       faster perceptron 0
-                   pairs 'XYX'
-                   hi lo 0.1
-          pair threshold 0
-          select weights 'last'
-                  l1 reg 0 'none'
-                    pclr no
-               max pairs 4294967295
-                  repeat 1
-                cdec cfg 'cdec.ini'
-                   input ''
+            error margin 1
+                  l1 reg 0
+            decoder conf 'cdec.ini'
+                   input 'work/shard.1.0.in'
                   output 'work/weights.1.0'
-(a dot represents 10 inputs)
+(a dot per input)
 Iteration #1 of 1.
-  3
+ .... 3
 WEIGHTS
-              Glue = -0.3229
-       WordPenalty = +0.27969
-     LanguageModel = +1.3645
- LanguageModel_OOV = -0.0443
-     PhraseModel_0 = -0.19049
-     PhraseModel_1 = -0.077698
-     PhraseModel_2 = +0.058898
-     PhraseModel_3 = +0.017251
-     PhraseModel_4 = -1.5474
-     PhraseModel_5 = +0
-     PhraseModel_6 = -0.1818
-       PassThrough = -0.193
+              Glue = -0.2722
+       WordPenalty = +0.05433
+     LanguageModel = +0.69948
+ LanguageModel_OOV = -0.2641
+     PhraseModel_0 = -1.4208
+     PhraseModel_1 = -1.563
+     PhraseModel_2 = -0.21051
+     PhraseModel_3 = -0.17764
+     PhraseModel_4 = -1.6583
+     PhraseModel_5 = +0.0794
+     PhraseModel_6 = +0.1528
+       PassThrough = -0.2367
         ---
-       1best avg score: 0.070229 (+0.070229)
- 1best avg model score: -44.01 (-44.01)
-           avg # pairs: 1294
-        avg # rank err: 878.67
-     avg # margin viol: 350.67
-       k-best loss imp: 100%
-    non0 feature count: 11
+       1best avg score: 0.071329 (+0.071329)
+ 1best avg model score: -41.362
+           avg # pairs: 1862.3
+   non-0 feature count: 12
            avg list sz: 100
-           avg f count: 11.487
-(time 0.28 min, 5.7 s/S)
-
-Writing weights file to 'work/weights.1.0' ...
-done
+           avg f count: 11.847
+(time 0.28 min, 5 s/S)
 
 ---
-Best iteration: 1 [SCORE 'stupid_bleu'=0.070229].
+Best iteration: 1 [GOLD = 0.071329].
 This took 0.28333 min.
diff --git a/training/dtrain/examples/parallelized/work/out.1.1 b/training/dtrain/examples/parallelized/work/out.1.1
index 9346fc82..79ac35dc 100644
--- a/training/dtrain/examples/parallelized/work/out.1.1
+++ b/training/dtrain/examples/parallelized/work/out.1.1
@@ -1,66 +1,44 @@
-                cdec cfg 'cdec.ini'
 Loading the LM will be faster if you build a binary file.
 Reading ../standard/nc-wmt11.en.srilm.gz
 ----5---10---15---20---25---30---35---40---45---50---55---60---65---70---75---80---85---90---95--100
 ****************************************************************************************************
-Seeding random number sequence to 3557309480
-
 dtrain
 Parameters:
                        k 100
                        N 4
                        T 1
-                   batch 0
-                  scorer 'stupid_bleu'
-             sample from 'kbest'
-                  filter 'uniq'
            learning rate 0.0001
-                   gamma 0
-             loss margin 1
-       faster perceptron 0
-                   pairs 'XYX'
-                   hi lo 0.1
-          pair threshold 0
-          select weights 'last'
-                  l1 reg 0 'none'
-                    pclr no
-               max pairs 4294967295
-                  repeat 1
-                cdec cfg 'cdec.ini'
-                   input ''
+            error margin 1
+                  l1 reg 0
+            decoder conf 'cdec.ini'
+                   input 'work/shard.1.0.in'
                   output 'work/weights.1.1'
               weights in 'work/weights.0'
-(a dot represents 10 inputs)
+(a dot per input)
 Iteration #1 of 1.
-  3
+ .... 3
 WEIGHTS
-              Glue = -0.26425
-       WordPenalty = +0.047881
-     LanguageModel = +0.78496
- LanguageModel_OOV = -0.49307
-     PhraseModel_0 = -0.58703
-     PhraseModel_1 = -0.33425
-     PhraseModel_2 = +0.20834
-     PhraseModel_3 = -0.043346
-     PhraseModel_4 = -0.60761
-     PhraseModel_5 = +0.123
-     PhraseModel_6 = -0.05415
-       PassThrough = -0.42167
+              Glue = -0.20488
+       WordPenalty = -0.0091745
+     LanguageModel = +0.79433
+ LanguageModel_OOV = -0.4309
+     PhraseModel_0 = -0.56242
+     PhraseModel_1 = +0.85363
+     PhraseModel_2 = +1.3458
+     PhraseModel_3 = -0.13095
+     PhraseModel_4 = -0.94762
+     PhraseModel_5 = +0.02435
+     PhraseModel_6 = -0.16003
+       PassThrough = -0.46105
         ---
-       1best avg score: 0.085952 (+0.085952)
- 1best avg model score: -45.175 (-45.175)
-           avg # pairs: 1180.7
-        avg # rank err: 668.33
-     avg # margin viol: 512.33
-       k-best loss imp: 100%
-    non0 feature count: 12
+       1best avg score: 0.13017 (+0.13017)
+ 1best avg model score: 14.53
+           avg # pairs: 1968
+   non-0 feature count: 12
            avg list sz: 100
-           avg f count: 12
-(time 0.27 min, 5.3 s/S)
-
-Writing weights file to 'work/weights.1.1' ...
-done
+           avg f count: 11
+(time 0.33 min, 6 s/S)
 
 ---
-Best iteration: 1 [SCORE 'stupid_bleu'=0.085952].
-This took 0.26667 min.
+Best iteration: 1 [GOLD = 0.13017].
+This took 0.33333 min.
diff --git a/training/dtrain/examples/parallelized/work/out.1.2 b/training/dtrain/examples/parallelized/work/out.1.2
index 08f07a75..8c4f8b03 100644
--- a/training/dtrain/examples/parallelized/work/out.1.2
+++ b/training/dtrain/examples/parallelized/work/out.1.2
@@ -1,66 +1,44 @@
-                cdec cfg 'cdec.ini'
 Loading the LM will be faster if you build a binary file.
 Reading ../standard/nc-wmt11.en.srilm.gz
 ----5---10---15---20---25---30---35---40---45---50---55---60---65---70---75---80---85---90---95--100
 ****************************************************************************************************
-Seeding random number sequence to 56743915
-
 dtrain
 Parameters:
                        k 100
                        N 4
                        T 1
-                   batch 0
-                  scorer 'stupid_bleu'
-             sample from 'kbest'
-                  filter 'uniq'
            learning rate 0.0001
-                   gamma 0
-             loss margin 1
-       faster perceptron 0
-                   pairs 'XYX'
-                   hi lo 0.1
-          pair threshold 0
-          select weights 'last'
-                  l1 reg 0 'none'
-                    pclr no
-               max pairs 4294967295
-                  repeat 1
-                cdec cfg 'cdec.ini'
-                   input ''
+            error margin 1
+                  l1 reg 0
+            decoder conf 'cdec.ini'
+                   input 'work/shard.1.0.in'
                   output 'work/weights.1.2'
               weights in 'work/weights.1'
-(a dot represents 10 inputs)
+(a dot per input)
 Iteration #1 of 1.
-  3
+ .... 3
 WEIGHTS
-              Glue = -0.23608
-       WordPenalty = +0.10931
-     LanguageModel = +0.81339
- LanguageModel_OOV = -0.33238
-     PhraseModel_0 = -0.53685
-     PhraseModel_1 = -0.049658
-     PhraseModel_2 = +0.40277
-     PhraseModel_3 = +0.14601
-     PhraseModel_4 = -0.72851
-     PhraseModel_5 = +0.03475
-     PhraseModel_6 = -0.27192
-       PassThrough = -0.34763
+              Glue = -0.49853
+       WordPenalty = +0.07636
+     LanguageModel = +1.3183
+ LanguageModel_OOV = -0.60902
+     PhraseModel_0 = -0.22481
+     PhraseModel_1 = +0.86369
+     PhraseModel_2 = +1.0747
+     PhraseModel_3 = +0.18002
+     PhraseModel_4 = -0.84661
+     PhraseModel_5 = +0.02435
+     PhraseModel_6 = +0.11247
+       PassThrough = -0.63918
         ---
-       1best avg score: 0.10073 (+0.10073)
- 1best avg model score: -38.422 (-38.422)
-           avg # pairs: 1505.3
-        avg # rank err: 777
-     avg # margin viol: 691.67
-       k-best loss imp: 100%
-    non0 feature count: 12
+       1best avg score: 0.15478 (+0.15478)
+ 1best avg model score: -7.2154
+           avg # pairs: 1776
+   non-0 feature count: 12
            avg list sz: 100
-           avg f count: 12
-(time 0.35 min, 7 s/S)
-
-Writing weights file to 'work/weights.1.2' ...
-done
+           avg f count: 11.327
+(time 0.27 min, 5 s/S)
 
 ---
-Best iteration: 1 [SCORE 'stupid_bleu'=0.10073].
-This took 0.35 min.
+Best iteration: 1 [GOLD = 0.15478].
+This took 0.26667 min.
diff --git a/training/dtrain/examples/parallelized/work/out.2.0 b/training/dtrain/examples/parallelized/work/out.2.0
index 25ef6d4e..07c85963 100644
--- a/training/dtrain/examples/parallelized/work/out.2.0
+++ b/training/dtrain/examples/parallelized/work/out.2.0
@@ -1,65 +1,43 @@
-                cdec cfg 'cdec.ini'
 Loading the LM will be faster if you build a binary file.
 Reading ../standard/nc-wmt11.en.srilm.gz
 ----5---10---15---20---25---30---35---40---45---50---55---60---65---70---75---80---85---90---95--100
 ****************************************************************************************************
-Seeding random number sequence to 2662215673
-
 dtrain
 Parameters:
                        k 100
                        N 4
                        T 1
-                   batch 0
-                  scorer 'stupid_bleu'
-             sample from 'kbest'
-                  filter 'uniq'
            learning rate 0.0001
-                   gamma 0
-             loss margin 1
-       faster perceptron 0
-                   pairs 'XYX'
-                   hi lo 0.1
-          pair threshold 0
-          select weights 'last'
-                  l1 reg 0 'none'
-                    pclr no
-               max pairs 4294967295
-                  repeat 1
-                cdec cfg 'cdec.ini'
-                   input ''
+            error margin 1
+                  l1 reg 0
+            decoder conf 'cdec.ini'
+                   input 'work/shard.2.0.in'
                   output 'work/weights.2.0'
-(a dot represents 10 inputs)
+(a dot per input)
 Iteration #1 of 1.
-  3
+ .... 3
 WEIGHTS
-              Glue = -0.1259
-       WordPenalty = +0.048294
-     LanguageModel = +0.36254
- LanguageModel_OOV = -0.1228
-     PhraseModel_0 = +0.26357
-     PhraseModel_1 = +0.24793
-     PhraseModel_2 = +0.0063763
-     PhraseModel_3 = -0.18966
-     PhraseModel_4 = -0.226
+              Glue = -0.2109
+       WordPenalty = +0.14922
+     LanguageModel = +0.79686
+ LanguageModel_OOV = -0.6627
+     PhraseModel_0 = +0.37999
+     PhraseModel_1 = +0.69213
+     PhraseModel_2 = +0.3422
+     PhraseModel_3 = +1.1426
+     PhraseModel_4 = -0.55413
      PhraseModel_5 = +0
-     PhraseModel_6 = +0.0743
-       PassThrough = -0.1335
+     PhraseModel_6 = +0.0676
+       PassThrough = -0.6343
         ---
-       1best avg score: 0.072836 (+0.072836)
- 1best avg model score: -0.56296 (-0.56296)
-           avg # pairs: 1094.7
-        avg # rank err: 658
-     avg # margin viol: 436.67
-       k-best loss imp: 100%
-    non0 feature count: 11
+       1best avg score: 0.072374 (+0.072374)
+ 1best avg model score: -27.384
+           avg # pairs: 2582
+   non-0 feature count: 11
            avg list sz: 100
-           avg f count: 10.813
-(time 0.13 min, 2.7 s/S)
-
-Writing weights file to 'work/weights.2.0' ...
-done
+           avg f count: 11.54
+(time 0.32 min, 6 s/S)
 
 ---
-Best iteration: 1 [SCORE 'stupid_bleu'=0.072836].
-This took 0.13333 min.
+Best iteration: 1 [GOLD = 0.072374].
+This took 0.31667 min.
diff --git a/training/dtrain/examples/parallelized/work/out.2.1 b/training/dtrain/examples/parallelized/work/out.2.1
index 8e4efde9..c54bb1b1 100644
--- a/training/dtrain/examples/parallelized/work/out.2.1
+++ b/training/dtrain/examples/parallelized/work/out.2.1
@@ -1,66 +1,44 @@
-                cdec cfg 'cdec.ini'
 Loading the LM will be faster if you build a binary file.
 Reading ../standard/nc-wmt11.en.srilm.gz
 ----5---10---15---20---25---30---35---40---45---50---55---60---65---70---75---80---85---90---95--100
 ****************************************************************************************************
-Seeding random number sequence to 3092904479
-
 dtrain
 Parameters:
                        k 100
                        N 4
                        T 1
-                   batch 0
-                  scorer 'stupid_bleu'
-             sample from 'kbest'
-                  filter 'uniq'
            learning rate 0.0001
-                   gamma 0
-             loss margin 1
-       faster perceptron 0
-                   pairs 'XYX'
-                   hi lo 0.1
-          pair threshold 0
-          select weights 'last'
-                  l1 reg 0 'none'
-                    pclr no
-               max pairs 4294967295
-                  repeat 1
-                cdec cfg 'cdec.ini'
-                   input ''
+            error margin 1
+                  l1 reg 0
+            decoder conf 'cdec.ini'
+                   input 'work/shard.2.0.in'
                   output 'work/weights.2.1'
               weights in 'work/weights.0'
-(a dot represents 10 inputs)
+(a dot per input)
 Iteration #1 of 1.
-  3
+ .... 3
 WEIGHTS
-              Glue = -0.10385
-       WordPenalty = +0.038717
-     LanguageModel = +0.49413
- LanguageModel_OOV = -0.24887
-     PhraseModel_0 = -0.32102
-     PhraseModel_1 = +0.34413
-     PhraseModel_2 = +0.62366
-     PhraseModel_3 = -0.49337
-     PhraseModel_4 = -0.77005
-     PhraseModel_5 = +0.007
-     PhraseModel_6 = -0.05055
-       PassThrough = -0.23928
+              Glue = -0.76608
+       WordPenalty = +0.15938
+     LanguageModel = +1.5897
+ LanguageModel_OOV = -0.521
+     PhraseModel_0 = -0.58348
+     PhraseModel_1 = +0.29828
+     PhraseModel_2 = +0.78493
+     PhraseModel_3 = +0.083222
+     PhraseModel_4 = -0.93843
+     PhraseModel_5 = +0.02435
+     PhraseModel_6 = -0.27382
+       PassThrough = -0.55115
         ---
-       1best avg score: 0.10245 (+0.10245)
- 1best avg model score: -20.384 (-20.384)
-           avg # pairs: 1741.7
-        avg # rank err: 953.67
-     avg # margin viol: 585.33
-       k-best loss imp: 100%
-    non0 feature count: 12
+       1best avg score: 0.12881 (+0.12881)
+ 1best avg model score: -9.6731
+           avg # pairs: 2020.3
+   non-0 feature count: 12
            avg list sz: 100
-           avg f count: 11.977
-(time 0.12 min, 2.3 s/S)
-
-Writing weights file to 'work/weights.2.1' ...
-done
+           avg f count: 12
+(time 0.32 min, 6 s/S)
 
 ---
-Best iteration: 1 [SCORE 'stupid_bleu'=0.10245].
-This took 0.11667 min.
+Best iteration: 1 [GOLD = 0.12881].
+This took 0.31667 min.
diff --git a/training/dtrain/examples/parallelized/work/out.2.2 b/training/dtrain/examples/parallelized/work/out.2.2
index e0ca2110..f5d6229f 100644
--- a/training/dtrain/examples/parallelized/work/out.2.2
+++ b/training/dtrain/examples/parallelized/work/out.2.2
@@ -1,66 +1,44 @@
-                cdec cfg 'cdec.ini'
 Loading the LM will be faster if you build a binary file.
 Reading ../standard/nc-wmt11.en.srilm.gz
 ----5---10---15---20---25---30---35---40---45---50---55---60---65---70---75---80---85---90---95--100
 ****************************************************************************************************
-Seeding random number sequence to 2803362953
-
 dtrain
 Parameters:
                        k 100
                        N 4
                        T 1
-                   batch 0
-                  scorer 'stupid_bleu'
-             sample from 'kbest'
-                  filter 'uniq'
            learning rate 0.0001
-                   gamma 0
-             loss margin 1
-       faster perceptron 0
-                   pairs 'XYX'
-                   hi lo 0.1
-          pair threshold 0
-          select weights 'last'
-                  l1 reg 0 'none'
-                    pclr no
-               max pairs 4294967295
-                  repeat 1
-                cdec cfg 'cdec.ini'
-                   input ''
+            error margin 1
+                  l1 reg 0
+            decoder conf 'cdec.ini'
+                   input 'work/shard.2.0.in'
                   output 'work/weights.2.2'
               weights in 'work/weights.1'
-(a dot represents 10 inputs)
+(a dot per input)
 Iteration #1 of 1.
-  3
+ .... 3
 WEIGHTS
-              Glue = -0.32907
-       WordPenalty = +0.049596
-     LanguageModel = +0.33496
- LanguageModel_OOV = -0.44357
-     PhraseModel_0 = -0.3068
-     PhraseModel_1 = +0.59376
-     PhraseModel_2 = +0.86416
-     PhraseModel_3 = -0.21072
-     PhraseModel_4 = -0.65734
-     PhraseModel_5 = +0.03475
-     PhraseModel_6 = -0.10653
-       PassThrough = -0.46082
+              Glue = -0.90863
+       WordPenalty = +0.10819
+     LanguageModel = +0.5239
+ LanguageModel_OOV = -0.41623
+     PhraseModel_0 = -0.86868
+     PhraseModel_1 = +0.40784
+     PhraseModel_2 = +1.1793
+     PhraseModel_3 = -0.24698
+     PhraseModel_4 = -1.2353
+     PhraseModel_5 = +0.03375
+     PhraseModel_6 = -0.17883
+       PassThrough = -0.44638
         ---
-       1best avg score: 0.25055 (+0.25055)
- 1best avg model score: -1.4459 (-1.4459)
-           avg # pairs: 1689
-        avg # rank err: 755.67
-     avg # margin viol: 829.33
-       k-best loss imp: 100%
-    non0 feature count: 12
+       1best avg score: 0.12788 (+0.12788)
+ 1best avg model score: 41.302
+           avg # pairs: 2246.3
+   non-0 feature count: 12
            avg list sz: 100
-           avg f count: 10.53
-(time 0.13 min, 2.7 s/S)
-
-Writing weights file to 'work/weights.2.2' ...
-done
+           avg f count: 10.98
+(time 0.35 min, 7 s/S)
 
 ---
-Best iteration: 1 [SCORE 'stupid_bleu'=0.25055].
-This took 0.13333 min.
+Best iteration: 1 [GOLD = 0.12788].
+This took 0.35 min.
diff --git a/training/dtrain/examples/parallelized/work/out.3.0 b/training/dtrain/examples/parallelized/work/out.3.0
index 3c074f04..fa499523 100644
--- a/training/dtrain/examples/parallelized/work/out.3.0
+++ b/training/dtrain/examples/parallelized/work/out.3.0
@@ -1,65 +1,43 @@
-                cdec cfg 'cdec.ini'
 Loading the LM will be faster if you build a binary file.
 Reading ../standard/nc-wmt11.en.srilm.gz
 ----5---10---15---20---25---30---35---40---45---50---55---60---65---70---75---80---85---90---95--100
 ****************************************************************************************************
-Seeding random number sequence to 316107185
-
 dtrain
 Parameters:
                        k 100
                        N 4
                        T 1
-                   batch 0
-                  scorer 'stupid_bleu'
-             sample from 'kbest'
-                  filter 'uniq'
            learning rate 0.0001
-                   gamma 0
-             loss margin 1
-       faster perceptron 0
-                   pairs 'XYX'
-                   hi lo 0.1
-          pair threshold 0
-          select weights 'last'
-                  l1 reg 0 'none'
-                    pclr no
-               max pairs 4294967295
-                  repeat 1
-                cdec cfg 'cdec.ini'
-                   input ''
+            error margin 1
+                  l1 reg 0
+            decoder conf 'cdec.ini'
+                   input 'work/shard.3.0.in'
                   output 'work/weights.3.0'
-(a dot represents 10 inputs)
+(a dot per input)
 Iteration #1 of 1.
-  1
+ .. 1
 WEIGHTS
-              Glue = +0.046
-       WordPenalty = +0.17328
-     LanguageModel = +1.1667
- LanguageModel_OOV = +0.066
-     PhraseModel_0 = -1.1694
-     PhraseModel_1 = -0.9883
-     PhraseModel_2 = +0.036205
-     PhraseModel_3 = -0.77387
-     PhraseModel_4 = -1.5019
-     PhraseModel_5 = +0.024
-     PhraseModel_6 = -0.514
-       PassThrough = +0.031
+              Glue = -0.09
+       WordPenalty = +0.32442
+     LanguageModel = +2.5769
+ LanguageModel_OOV = -0.009
+     PhraseModel_0 = -0.58972
+     PhraseModel_1 = +0.063691
+     PhraseModel_2 = +0.5366
+     PhraseModel_3 = +0.12867
+     PhraseModel_4 = -1.9801
+     PhraseModel_5 = +0.018
+     PhraseModel_6 = -0.486
+       PassThrough = -0.09
         ---
-       1best avg score: 0.032916 (+0.032916)
- 1best avg model score: 0 (+0)
-           avg # pairs: 900
-        avg # rank err: 900
-     avg # margin viol: 0
-       k-best loss imp: 100%
-    non0 feature count: 12
+       1best avg score: 0.034204 (+0.034204)
+ 1best avg model score: 0
+           avg # pairs: 1700
+   non-0 feature count: 12
            avg list sz: 100
-           avg f count: 11.72
-(time 0.23 min, 14 s/S)
-
-Writing weights file to 'work/weights.3.0' ...
-done
+           avg f count: 10.8
+(time 0.1 min, 6 s/S)
 
 ---
-Best iteration: 1 [SCORE 'stupid_bleu'=0.032916].
-This took 0.23333 min.
+Best iteration: 1 [GOLD = 0.034204].
+This took 0.1 min.
diff --git a/training/dtrain/examples/parallelized/work/out.3.1 b/training/dtrain/examples/parallelized/work/out.3.1
index 241d3455..c4b3aa3c 100644
--- a/training/dtrain/examples/parallelized/work/out.3.1
+++ b/training/dtrain/examples/parallelized/work/out.3.1
@@ -1,66 +1,44 @@
-                cdec cfg 'cdec.ini'
 Loading the LM will be faster if you build a binary file.
 Reading ../standard/nc-wmt11.en.srilm.gz
 ----5---10---15---20---25---30---35---40---45---50---55---60---65---70---75---80---85---90---95--100
 ****************************************************************************************************
-Seeding random number sequence to 353677750
-
 dtrain
 Parameters:
                        k 100
                        N 4
                        T 1
-                   batch 0
-                  scorer 'stupid_bleu'
-             sample from 'kbest'
-                  filter 'uniq'
            learning rate 0.0001
-                   gamma 0
-             loss margin 1
-       faster perceptron 0
-                   pairs 'XYX'
-                   hi lo 0.1
-          pair threshold 0
-          select weights 'last'
-                  l1 reg 0 'none'
-                    pclr no
-               max pairs 4294967295
-                  repeat 1
-                cdec cfg 'cdec.ini'
-                   input ''
+            error margin 1
+                  l1 reg 0
+            decoder conf 'cdec.ini'
+                   input 'work/shard.3.0.in'
                   output 'work/weights.3.1'
               weights in 'work/weights.0'
-(a dot represents 10 inputs)
+(a dot per input)
 Iteration #1 of 1.
-  1
+ .. 1
 WEIGHTS
-              Glue = -0.08475
-       WordPenalty = +0.11151
-     LanguageModel = +1.0635
- LanguageModel_OOV = -0.11468
-     PhraseModel_0 = -0.062922
-     PhraseModel_1 = +0.0035552
-     PhraseModel_2 = +0.039692
-     PhraseModel_3 = +0.080265
-     PhraseModel_4 = -0.57787
-     PhraseModel_5 = +0.0174
-     PhraseModel_6 = -0.17095
-       PassThrough = -0.18248
+              Glue = +0.31832
+       WordPenalty = +0.11139
+     LanguageModel = +0.95438
+ LanguageModel_OOV = -0.0608
+     PhraseModel_0 = -0.98113
+     PhraseModel_1 = -0.090531
+     PhraseModel_2 = +0.79088
+     PhraseModel_3 = -0.57623
+     PhraseModel_4 = -1.4382
+     PhraseModel_5 = +0.02435
+     PhraseModel_6 = -0.10812
+       PassThrough = -0.09095
         ---
-       1best avg score: 0.16117 (+0.16117)
- 1best avg model score: -67.89 (-67.89)
-           avg # pairs: 1411
-        avg # rank err: 460
-     avg # margin viol: 951
-       k-best loss imp: 100%
-    non0 feature count: 12
+       1best avg score: 0.084989 (+0.084989)
+ 1best avg model score: -52.323
+           avg # pairs: 2487
+   non-0 feature count: 12
            avg list sz: 100
            avg f count: 12
-(time 0.22 min, 13 s/S)
-
-Writing weights file to 'work/weights.3.1' ...
-done
+(time 0.1 min, 6 s/S)
 
 ---
-Best iteration: 1 [SCORE 'stupid_bleu'=0.16117].
-This took 0.21667 min.
+Best iteration: 1 [GOLD = 0.084989].
+This took 0.1 min.
diff --git a/training/dtrain/examples/parallelized/work/out.3.2 b/training/dtrain/examples/parallelized/work/out.3.2
index b995daf5..eb27dac2 100644
--- a/training/dtrain/examples/parallelized/work/out.3.2
+++ b/training/dtrain/examples/parallelized/work/out.3.2
@@ -1,66 +1,44 @@
-                cdec cfg 'cdec.ini'
 Loading the LM will be faster if you build a binary file.
 Reading ../standard/nc-wmt11.en.srilm.gz
 ----5---10---15---20---25---30---35---40---45---50---55---60---65---70---75---80---85---90---95--100
 ****************************************************************************************************
-Seeding random number sequence to 3001145976
-
 dtrain
 Parameters:
                        k 100
                        N 4
                        T 1
-                   batch 0
-                  scorer 'stupid_bleu'
-             sample from 'kbest'
-                  filter 'uniq'
            learning rate 0.0001
-                   gamma 0
-             loss margin 1
-       faster perceptron 0
-                   pairs 'XYX'
-                   hi lo 0.1
-          pair threshold 0
-          select weights 'last'
-                  l1 reg 0 'none'
-                    pclr no
-               max pairs 4294967295
-                  repeat 1
-                cdec cfg 'cdec.ini'
-                   input ''
+            error margin 1
+                  l1 reg 0
+            decoder conf 'cdec.ini'
+                   input 'work/shard.3.0.in'
                   output 'work/weights.3.2'
               weights in 'work/weights.1'
-(a dot represents 10 inputs)
+(a dot per input)
 Iteration #1 of 1.
-  1
+ .. 1
 WEIGHTS
-              Glue = -0.13247
-       WordPenalty = +0.053592
-     LanguageModel = +0.72105
- LanguageModel_OOV = -0.30827
-     PhraseModel_0 = -0.37053
-     PhraseModel_1 = +0.17551
-     PhraseModel_2 = +0.5
-     PhraseModel_3 = -0.1459
-     PhraseModel_4 = -0.59563
-     PhraseModel_5 = +0.03475
-     PhraseModel_6 = -0.11143
-       PassThrough = -0.32553
+              Glue = -0.12993
+       WordPenalty = +0.13651
+     LanguageModel = +0.58946
+ LanguageModel_OOV = -0.48362
+     PhraseModel_0 = -0.81262
+     PhraseModel_1 = +0.44273
+     PhraseModel_2 = +1.1733
+     PhraseModel_3 = -0.1826
+     PhraseModel_4 = -1.2213
+     PhraseModel_5 = +0.02435
+     PhraseModel_6 = -0.18823
+       PassThrough = -0.51378
         ---
-       1best avg score: 0.12501 (+0.12501)
- 1best avg model score: -62.128 (-62.128)
-           avg # pairs: 979
-        avg # rank err: 539
-     avg # margin viol: 440
-       k-best loss imp: 100%
-    non0 feature count: 12
+       1best avg score: 0.12674 (+0.12674)
+ 1best avg model score: -7.2878
+           avg # pairs: 1769
+   non-0 feature count: 12
            avg list sz: 100
            avg f count: 12
-(time 0.22 min, 13 s/S)
-
-Writing weights file to 'work/weights.3.2' ...
-done
+(time 0.1 min, 6 s/S)
 
 ---
-Best iteration: 1 [SCORE 'stupid_bleu'=0.12501].
-This took 0.21667 min.
+Best iteration: 1 [GOLD = 0.12674].
+This took 0.1 min.
diff --git a/training/dtrain/examples/parallelized/work/shard.0.0.in b/training/dtrain/examples/parallelized/work/shard.0.0.in
index d1b48321..a0ef6f54 100644
--- a/training/dtrain/examples/parallelized/work/shard.0.0.in
+++ b/training/dtrain/examples/parallelized/work/shard.0.0.in
@@ -1,3 +1,3 @@
-<seg grammar="grammar/grammar.out.8.gz" id="8">der erste schritt , um mit der rassenfrage umzugehen ist , ursache und folgen rassistischer feindseligkeiten zu verstehen , auch dann , wenn das bedeutet , unangenehme tatsachen aufzudecken .</seg> ||| the first step to address racial politics is to understand the origin and consequences of racial animosity , even if it means uncovering unpleasant truths .
-<seg grammar="grammar/grammar.out.5.gz" id="5">die großen parteien der rechten und der linken mitte haben sich dem problem gestellt , in dem sie den kopf in den sand gesteckt und allen aussichten zuwider gehofft haben , es möge bald verschwinden .</seg> ||| mainstream parties of the center left and center right have confronted this prospect by hiding their heads in the ground , hoping against hope that the problem will disappear .
+<seg grammar="grammar/grammar.out.0.gz" id="0">europas nach rassen geteiltes haus</seg> ||| europe 's divided racial house
+<seg grammar="grammar/grammar.out.1.gz" id="1">ein gemeinsames merkmal aller extremen rechten in europa ist ihr rassismus und die tatsache , daß sie das einwanderungsproblem als politischen hebel benutzen .</seg> ||| a common feature of europe 's extreme right is its racism and use of the immigration issue as a political wedge .
 <seg grammar="grammar/grammar.out.2.gz" id="2">der lega nord in italien , der vlaams block in den niederlanden , die anhänger von le pens nationaler front in frankreich , sind beispiele für parteien oder bewegungen , die sich um das gemeinsame thema : ablehnung der zuwanderung gebildet haben und um forderung nach einer vereinfachten politik , um sie zu regeln .</seg> ||| the lega nord in italy , the vlaams blok in the netherlands , the supporters of le pen 's national front in france , are all examples of parties or movements formed on the common theme of aversion to immigrants and promotion of simplistic policies to control them .
diff --git a/training/dtrain/examples/parallelized/work/shard.1.0.in b/training/dtrain/examples/parallelized/work/shard.1.0.in
index a63f05bd..05f0273b 100644
--- a/training/dtrain/examples/parallelized/work/shard.1.0.in
+++ b/training/dtrain/examples/parallelized/work/shard.1.0.in
@@ -1,3 +1,3 @@
-<seg grammar="grammar/grammar.out.4.gz" id="4">eine alternde einheimische bevölkerung und immer offenere grenzen vermehren die rassistische zersplitterung in den europäischen ländern .</seg> ||| an aging population at home and ever more open borders imply increasing racial fragmentation in european countries .
-<seg grammar="grammar/grammar.out.9.gz" id="9">genau das haben in den usa eine große anzahl an forschungsvorhaben in wirtschaft , soziologie , psychologie und politikwissenschaft geleistet . diese forschungen zeigten , daß menschen unterschiedlicher rasse einander deutlich weniger vertrauen .</seg> ||| this is precisely what a large amount of research in economics , sociology , psychology and political science has done for the us .
 <seg grammar="grammar/grammar.out.3.gz" id="3">während individuen wie jörg haidar und jean @-@ marie le pen kommen und ( leider nicht zu bald ) wieder gehen mögen , wird die rassenfrage aus der europäischer politik nicht so bald verschwinden .</seg> ||| while individuals like jorg haidar and jean @-@ marie le pen may come and ( never to soon ) go , the race question will not disappear from european politics anytime soon .
+<seg grammar="grammar/grammar.out.4.gz" id="4">eine alternde einheimische bevölkerung und immer offenere grenzen vermehren die rassistische zersplitterung in den europäischen ländern .</seg> ||| an aging population at home and ever more open borders imply increasing racial fragmentation in european countries .
+<seg grammar="grammar/grammar.out.5.gz" id="5">die großen parteien der rechten und der linken mitte haben sich dem problem gestellt , in dem sie den kopf in den sand gesteckt und allen aussichten zuwider gehofft haben , es möge bald verschwinden .</seg> ||| mainstream parties of the center left and center right have confronted this prospect by hiding their heads in the ground , hoping against hope that the problem will disappear .
diff --git a/training/dtrain/examples/parallelized/work/shard.2.0.in b/training/dtrain/examples/parallelized/work/shard.2.0.in
index fe542b40..0528d357 100644
--- a/training/dtrain/examples/parallelized/work/shard.2.0.in
+++ b/training/dtrain/examples/parallelized/work/shard.2.0.in
@@ -1,3 +1,3 @@
-<seg grammar="grammar/grammar.out.1.gz" id="1">ein gemeinsames merkmal aller extremen rechten in europa ist ihr rassismus und die tatsache , daß sie das einwanderungsproblem als politischen hebel benutzen .</seg> ||| a common feature of europe 's extreme right is its racism and use of the immigration issue as a political wedge .
-<seg grammar="grammar/grammar.out.0.gz" id="0">europas nach rassen geteiltes haus</seg> ||| europe 's divided racial house
 <seg grammar="grammar/grammar.out.6.gz" id="6">das aber wird es nicht , wie die geschichte des rassismus in amerika deutlich zeigt .</seg> ||| it will not , as america 's racial history clearly shows .
+<seg grammar="grammar/grammar.out.7.gz" id="7">die beziehungen zwischen den rassen standen in den usa über jahrzehnte - und tun das noch heute - im zentrum der politischen debatte . das ging so weit , daß rassentrennung genauso wichtig wie das einkommen wurde , - wenn nicht sogar noch wichtiger - um politische zuneigungen und einstellungen zu bestimmen .</seg> ||| race relations in the us have been for decades - and remain - at the center of political debate , to the point that racial cleavages are as important as income , if not more , as determinants of political preferences and attitudes .
+<seg grammar="grammar/grammar.out.8.gz" id="8">der erste schritt , um mit der rassenfrage umzugehen ist , ursache und folgen rassistischer feindseligkeiten zu verstehen , auch dann , wenn das bedeutet , unangenehme tatsachen aufzudecken .</seg> ||| the first step to address racial politics is to understand the origin and consequences of racial animosity , even if it means uncovering unpleasant truths .
diff --git a/training/dtrain/examples/parallelized/work/shard.3.0.in b/training/dtrain/examples/parallelized/work/shard.3.0.in
index 4a8fa5b1..f7cbb3e3 100644
--- a/training/dtrain/examples/parallelized/work/shard.3.0.in
+++ b/training/dtrain/examples/parallelized/work/shard.3.0.in
@@ -1 +1 @@
-<seg grammar="grammar/grammar.out.7.gz" id="7">die beziehungen zwischen den rassen standen in den usa über jahrzehnte - und tun das noch heute - im zentrum der politischen debatte . das ging so weit , daß rassentrennung genauso wichtig wie das einkommen wurde , - wenn nicht sogar noch wichtiger - um politische zuneigungen und einstellungen zu bestimmen .</seg> ||| race relations in the us have been for decades - and remain - at the center of political debate , to the point that racial cleavages are as important as income , if not more , as determinants of political preferences and attitudes .
+<seg grammar="grammar/grammar.out.9.gz" id="9">genau das haben in den usa eine große anzahl an forschungsvorhaben in wirtschaft , soziologie , psychologie und politikwissenschaft geleistet . diese forschungen zeigten , daß menschen unterschiedlicher rasse einander deutlich weniger vertrauen .</seg> ||| this is precisely what a large amount of research in economics , sociology , psychology and political science has done for the us .
diff --git a/training/dtrain/examples/parallelized/work/weights.0 b/training/dtrain/examples/parallelized/work/weights.0
index c560fdbd..816269cd 100644
--- a/training/dtrain/examples/parallelized/work/weights.0
+++ b/training/dtrain/examples/parallelized/work/weights.0
@@ -1,12 +1,12 @@
-PhraseModel_4	-0.6990170657294328
-LanguageModel	0.891784887346263
-PhraseModel_0	-0.2107507586515428
-PhraseModel_1	-0.15442709655871997
-PhraseModel_3	-0.07262514338204715
-PhraseModel_6	-0.10965000000000148
-Glue	-0.03644999999999783
-WordPenalty	0.13204723722268177
-PassThrough	-0.09437500000000089
-LanguageModel_OOV	-0.036775000000000564
-PhraseModel_2	0.025521702385571707
-PhraseModel_5	0.006999999999999977
+LanguageModel	1.200704259340465
+PhraseModel_4	-1.2434381298299035
+PhraseModel_1	0.050697726409824076
+PhraseModel_0	-0.516923312932941
+PhraseModel_2	0.5051987092783867
+PhraseModel_3	0.20955092377784057
+PassThrough	-0.32285
+LanguageModel_OOV	-0.29269999999999996
+PhraseModel_6	-0.158425
+Glue	-0.05817500000000002
+WordPenalty	0.12758486142112804
+PhraseModel_5	0.02435
diff --git a/training/dtrain/examples/parallelized/work/weights.0.0 b/training/dtrain/examples/parallelized/work/weights.0.0
index 91eedc7b..be386c62 100644
--- a/training/dtrain/examples/parallelized/work/weights.0.0
+++ b/training/dtrain/examples/parallelized/work/weights.0.0
@@ -1,12 +1,11 @@
-PassThrough	-0.082000000000001058
-Glue	0.25700000000000267
-LanguageModel_OOV	-0.046000000000000034
-LanguageModel	0.67341721152744249
-PhraseModel_6	0.18290000000000028
-PhraseModel_5	0.0039999999999999975
-PhraseModel_4	0.47916377173928498
-PhraseModel_3	0.65577926367715722
-PhraseModel_2	0.00060731048591637909
-PhraseModel_0	0.25329462707903372
-WordPenalty	0.026926257878001431
-PhraseModel_1	0.20035945197369062
+WordPenalty -0.017632355965271129
+LanguageModel 0.72957628464102753
+LanguageModel_OOV -0.23499999999999999
+PhraseModel_0 -0.43720953659541578
+PhraseModel_1 1.0100170838129212
+PhraseModel_2 1.3524984123857073
+PhraseModel_3 -0.25541132249775761
+PhraseModel_4 -0.78115161368856911
+PhraseModel_6 -0.36810000000000004
+Glue 0.34040000000000004
+PassThrough -0.33040000000000003
diff --git a/training/dtrain/examples/parallelized/work/weights.0.1 b/training/dtrain/examples/parallelized/work/weights.0.1
index 6fcc9999..d4c77d07 100644
--- a/training/dtrain/examples/parallelized/work/weights.0.1
+++ b/training/dtrain/examples/parallelized/work/weights.0.1
@@ -1,12 +1,12 @@
-PassThrough	-0.2346750000000028
-Glue	-0.17904999999999763
-WordPenalty	0.062125825636256168
-LanguageModel	0.66824625053667575
-LanguageModel_OOV	-0.15247500000000355
-PhraseModel_0	-0.5581144363944085
-PhraseModel_1	0.12740874153205478
-PhraseModel_2	0.6038779278708799
-PhraseModel_3	-0.44463820299544454
-PhraseModel_4	-0.63136538282212662
-PhraseModel_5	-0.0084000000000000324
-PhraseModel_6	-0.20164999999999911
+WordPenalty 0.12966947493426365
+LanguageModel 0.3989224621154368
+LanguageModel_OOV -0.63139999999999996
+PhraseModel_0 -0.63991953012355962
+PhraseModel_1 0.74197897612368646
+PhraseModel_2 1.3096163833051435
+PhraseModel_3 -0.12160001974680773
+PhraseModel_4 -1.2274031286515816
+PhraseModel_5 0.02435
+PhraseModel_6 -0.210925
+Glue -0.40907500000000002
+PassThrough -0.66155000000000008
diff --git a/training/dtrain/examples/parallelized/work/weights.0.2 b/training/dtrain/examples/parallelized/work/weights.0.2
index 5668915d..8ce1449b 100644
--- a/training/dtrain/examples/parallelized/work/weights.0.2
+++ b/training/dtrain/examples/parallelized/work/weights.0.2
@@ -1,12 +1,12 @@
-PassThrough	-0.38122499999999337
-Glue	-0.019274999999998679
-WordPenalty	0.022192448025253487
-LanguageModel	0.4068780855136106
-LanguageModel_OOV	-0.363974999999992
-PhraseModel_0	-0.36273429313029715
-PhraseModel_1	0.56431752511029298
-PhraseModel_2	0.85638010019687694
-PhraseModel_3	-0.20222345248738063
-PhraseModel_4	-0.48295466434310252
-PhraseModel_5	0.031450000000000339
-PhraseModel_6	-0.26092499999998625
+WordPenalty 0.10319922626226019
+LanguageModel 0.6647396869692952
+LanguageModel_OOV -0.622525
+PhraseModel_0 -0.59993441316076157
+PhraseModel_1 0.78991513935858193
+PhraseModel_2 1.3148638774685031
+PhraseModel_3 0.2143393571820455
+PhraseModel_4 -1.0173894637028262
+PhraseModel_5 0.02435
+PhraseModel_6 -0.18452499999999999
+Glue -0.44422499999999998
+PassThrough -0.65267500000000012
diff --git a/training/dtrain/examples/parallelized/work/weights.1 b/training/dtrain/examples/parallelized/work/weights.1
index f52e07b8..2a00be2e 100644
--- a/training/dtrain/examples/parallelized/work/weights.1
+++ b/training/dtrain/examples/parallelized/work/weights.1
@@ -1,12 +1,12 @@
-LanguageModel	0.7527067666152598
-PhraseModel_4	-0.6467221787583058
-PhraseModel_2	0.36889175522051015
-PhraseModel_0	-0.38227173053779245
-PhraseModel_3	-0.2252732111174934
-LanguageModel_OOV	-0.25227499999999975
-PassThrough	-0.2695250000000011
-PhraseModel_1	0.03521067244127414
-Glue	-0.1579749999999981
-PhraseModel_6	-0.11932500000000047
-WordPenalty	0.0650573133891042
-PhraseModel_5	0.03475000000000043
+PhraseModel_4	-1.1379250444170055
+PhraseModel_2	1.0578050661336098
+LanguageModel	0.9343385461706668
+PhraseModel_0	-0.6917392152965985
+PhraseModel_1	0.4508371141128957
+PassThrough	-0.4411750000000001
+Glue	-0.265425
+LanguageModel_OOV	-0.411025
+PhraseModel_3	-0.186390082624459
+PhraseModel_6	-0.188225
+WordPenalty	0.09781397468665984
+PhraseModel_5	0.02435
diff --git a/training/dtrain/examples/parallelized/work/weights.1.0 b/training/dtrain/examples/parallelized/work/weights.1.0
index 31e08d81..cdcf959e 100644
--- a/training/dtrain/examples/parallelized/work/weights.1.0
+++ b/training/dtrain/examples/parallelized/work/weights.1.0
@@ -1,11 +1,12 @@
-LanguageModel_OOV	-0.044300000000000235
-PassThrough	-0.19300000000000087
-PhraseModel_6	-0.18180000000000701
-LanguageModel	1.3644969337716422
-PhraseModel_3	0.017250706134911725
-PhraseModel_4	-1.5473728273858063
-Glue	-0.32289999999999447
-PhraseModel_1	-0.077697953502182365
-WordPenalty	0.27968564634568688
-PhraseModel_0	-0.19048660891012237
-PhraseModel_2	0.05889844333199834
+WordPenalty 0.05433023968609621
+LanguageModel 0.69947965605855011
+LanguageModel_OOV -0.2641
+PhraseModel_0 -1.4207505705360111
+PhraseModel_1 -1.563047680441811
+PhraseModel_2 -0.21050528366541305
+PhraseModel_3 -0.17764037275860439
+PhraseModel_4 -1.6583462458159566
+PhraseModel_5 0.079399999999999998
+PhraseModel_6 0.15280000000000002
+Glue -0.27220000000000005
+PassThrough -0.23670000000000002
diff --git a/training/dtrain/examples/parallelized/work/weights.1.1 b/training/dtrain/examples/parallelized/work/weights.1.1
index 544ff462..c1bb2cf0 100644
--- a/training/dtrain/examples/parallelized/work/weights.1.1
+++ b/training/dtrain/examples/parallelized/work/weights.1.1
@@ -1,12 +1,12 @@
-PassThrough	-0.42167499999999858
-Glue	-0.26424999999999721
-WordPenalty	0.04788096662983269
-LanguageModel	0.78495517342352483
-LanguageModel_OOV	-0.49307499999999477
-PhraseModel_0	-0.58703462849498356
-PhraseModel_1	-0.33425278954714266
-PhraseModel_2	0.20834221229630179
-PhraseModel_3	-0.043345645640208569
-PhraseModel_4	-0.60760531115816907
-PhraseModel_5	0.12300000000000186
-PhraseModel_6	-0.054150000000001031
+WordPenalty -0.0091744709302067785
+LanguageModel 0.79433413663506514
+LanguageModel_OOV -0.43090000000000001
+PhraseModel_0 -0.56242499947237046
+PhraseModel_1 0.85362516703032698
+PhraseModel_2 1.3457900890481096
+PhraseModel_3 -0.13095079554478939
+PhraseModel_4 -0.94761908497413061
+PhraseModel_5 0.02435
+PhraseModel_6 -0.160025
+Glue -0.20487500000000003
+PassThrough -0.46105000000000007
diff --git a/training/dtrain/examples/parallelized/work/weights.1.2 b/training/dtrain/examples/parallelized/work/weights.1.2
index ac3284b9..c9598a04 100644
--- a/training/dtrain/examples/parallelized/work/weights.1.2
+++ b/training/dtrain/examples/parallelized/work/weights.1.2
@@ -1,12 +1,12 @@
-PassThrough	-0.34762500000000224
-Glue	-0.23607500000000026
-WordPenalty	0.10931192109504413
-LanguageModel	0.81339027211983694
-LanguageModel_OOV	-0.33237500000000098
-PhraseModel_0	-0.53685104648974269
-PhraseModel_1	-0.049657790506137042
-PhraseModel_2	0.40277066454544108
-PhraseModel_3	0.14600791389785803
-PhraseModel_4	-0.72850673041349101
-PhraseModel_5	0.034750000000000433
-PhraseModel_6	-0.27192499999999448
+WordPenalty 0.076359827280638559
+LanguageModel 1.3183380272921175
+LanguageModel_OOV -0.60902499999999993
+PhraseModel_0 -0.2248075206657828
+PhraseModel_1 0.86368802571834491
+PhraseModel_2 1.0746702462261808
+PhraseModel_3 0.18002263643876637
+PhraseModel_4 -0.84660750337519441
+PhraseModel_5 0.02435
+PhraseModel_6 0.11247499999999999
+Glue -0.49852500000000005
+PassThrough -0.63917500000000005
diff --git a/training/dtrain/examples/parallelized/work/weights.2 b/training/dtrain/examples/parallelized/work/weights.2
index dedaf165..310973ec 100644
--- a/training/dtrain/examples/parallelized/work/weights.2
+++ b/training/dtrain/examples/parallelized/work/weights.2
@@ -1,12 +1,12 @@
-PhraseModel_2	0.6558266927225778
-PhraseModel_4	-0.6161090299356294
-LanguageModel	0.5690697644415413
-PhraseModel_1	0.32098232482479416
-PhraseModel_0	-0.39422813904895143
-PassThrough	-0.37879999999999764
-LanguageModel_OOV	-0.3620499999999963
-Glue	-0.1792249999999967
-PhraseModel_6	-0.18769999999999526
-PhraseModel_3	-0.10321074877850786
-WordPenalty	0.05867318450512617
-PhraseModel_5	0.03392500000000041
+PhraseModel_2	1.185520780812669
+PhraseModel_4	-1.0801541070647134
+LanguageModel	0.7741099486587568
+PhraseModel_0	-0.6265095873268189
+PhraseModel_1	0.6260421233840029
+PassThrough	-0.5630000000000002
+Glue	-0.495325
+LanguageModel_OOV	-0.53285
+PhraseModel_3	-0.008805626854390465
+PhraseModel_6	-0.10977500000000001
+WordPenalty	0.1060655698428214
+PhraseModel_5	0.026699999999999998
diff --git a/training/dtrain/examples/parallelized/work/weights.2.0 b/training/dtrain/examples/parallelized/work/weights.2.0
index f7ece54d..3e87fed4 100644
--- a/training/dtrain/examples/parallelized/work/weights.2.0
+++ b/training/dtrain/examples/parallelized/work/weights.2.0
@@ -1,11 +1,11 @@
-LanguageModel_OOV	-0.12280000000000209
-PassThrough	-0.13350000000000165
-Glue	-0.1259000000000001
-PhraseModel_1	0.24792740418949952
-WordPenalty	0.048293546387642321
-PhraseModel_0	0.26356693580129958
-PhraseModel_2	0.0063762787517740458
-PhraseModel_3	-0.18966358382769741
-PhraseModel_4	-0.22599681869670471
-PhraseModel_6	0.074299999999999047
-LanguageModel	0.3625416478537038
+WordPenalty 0.14922358398195767
+LanguageModel 0.79685677298009394
+LanguageModel_OOV -0.66270000000000007
+PhraseModel_0 0.37998874905310187
+PhraseModel_1 0.69213063228111271
+PhraseModel_2 0.34219807728516061
+PhraseModel_3 1.1425846772648622
+PhraseModel_4 -0.55412548521619742
+PhraseModel_6 0.067599999999999993
+Glue -0.21090000000000003
+PassThrough -0.63429999999999997
diff --git a/training/dtrain/examples/parallelized/work/weights.2.1 b/training/dtrain/examples/parallelized/work/weights.2.1
index 0946609d..d129dc49 100644
--- a/training/dtrain/examples/parallelized/work/weights.2.1
+++ b/training/dtrain/examples/parallelized/work/weights.2.1
@@ -1,12 +1,12 @@
-PassThrough	-0.23927500000000015
-Glue	-0.10384999999999919
-WordPenalty	0.038717353061671053
-LanguageModel	0.49412782572695274
-LanguageModel_OOV	-0.24887499999999915
-PhraseModel_0	-0.32101572713801541
-PhraseModel_1	0.34413149733472631
-PhraseModel_2	0.62365535622061474
-PhraseModel_3	-0.49337445280658987
-PhraseModel_4	-0.77004673375347765
-PhraseModel_5	0.0069999999999999767
-PhraseModel_6	-0.05055000000000108
+WordPenalty 0.1593752174964457
+LanguageModel 1.5897162231676281
+LanguageModel_OOV -0.52100000000000002
+PhraseModel_0 -0.5834836741748588
+PhraseModel_1 0.29827543837280185
+PhraseModel_2 0.78493316593562568
+PhraseModel_3 0.083221832554333464
+PhraseModel_4 -0.93843312963279457
+PhraseModel_5 0.02435
+PhraseModel_6 -0.27382499999999999
+Glue -0.76607500000000006
+PassThrough -0.55115000000000003
diff --git a/training/dtrain/examples/parallelized/work/weights.2.2 b/training/dtrain/examples/parallelized/work/weights.2.2
index b766fc75..bcc83b44 100644
--- a/training/dtrain/examples/parallelized/work/weights.2.2
+++ b/training/dtrain/examples/parallelized/work/weights.2.2
@@ -1,12 +1,12 @@
-PassThrough	-0.46082499999999499
-Glue	-0.32907499999998979
-WordPenalty	0.049596429833348527
-LanguageModel	0.33496341201347335
-LanguageModel_OOV	-0.44357499999999361
-PhraseModel_0	-0.30679883980783829
-PhraseModel_1	0.5937585900939707
-PhraseModel_2	0.86415970329021152
-PhraseModel_3	-0.21072279838022553
-PhraseModel_4	-0.65734339854224544
-PhraseModel_5	0.034750000000000433
-PhraseModel_6	-0.10652500000000011
+WordPenalty 0.10819361280414735
+LanguageModel 0.52389743342585859
+LanguageModel_OOV -0.41622500000000001
+PhraseModel_0 -0.86867995703334211
+PhraseModel_1 0.40783818771767943
+PhraseModel_2 1.1792706530114188
+PhraseModel_3 -0.2469805689928464
+PhraseModel_4 -1.2352895858909159
+PhraseModel_5 0.033750000000000002
+PhraseModel_6 -0.17882500000000001
+Glue -0.90862500000000002
+PassThrough -0.44637500000000013
diff --git a/training/dtrain/examples/parallelized/work/weights.3.0 b/training/dtrain/examples/parallelized/work/weights.3.0
index 403ffbb3..e3586048 100644
--- a/training/dtrain/examples/parallelized/work/weights.3.0
+++ b/training/dtrain/examples/parallelized/work/weights.3.0
@@ -1,12 +1,12 @@
-PhraseModel_4	-1.501862388574505
-PhraseModel_3	-0.77386695951256013
-PhraseModel_6	-0.51399999999999824
-PhraseModel_5	0.02399999999999991
-LanguageModel	1.1666837562322641
-PhraseModel_2	0.036204776972598059
-PassThrough	0.030999999999999975
-Glue	0.046000000000000582
-PhraseModel_1	-0.98829728889588764
-WordPenalty	0.1732834982793964
-PhraseModel_0	-1.1693779885763822
-LanguageModel_OOV	0.066000000000000086
+WordPenalty 0.32441797798172944
+LanguageModel 2.5769043236821889
+LanguageModel_OOV -0.0090000000000000011
+PhraseModel_0 -0.58972189365343919
+PhraseModel_1 0.063690869987073351
+PhraseModel_2 0.53660363110809217
+PhraseModel_3 0.12867071310286207
+PhraseModel_4 -1.9801291745988916
+PhraseModel_5 0.018000000000000002
+PhraseModel_6 -0.48600000000000004
+Glue -0.090000000000000011
+PassThrough -0.090000000000000011
diff --git a/training/dtrain/examples/parallelized/work/weights.3.1 b/training/dtrain/examples/parallelized/work/weights.3.1
index c171d586..b27687d3 100644
--- a/training/dtrain/examples/parallelized/work/weights.3.1
+++ b/training/dtrain/examples/parallelized/work/weights.3.1
@@ -1,12 +1,12 @@
-PassThrough	-0.18247500000000313
-Glue	-0.084749999999998368
-WordPenalty	0.11150510822865688
-LanguageModel	1.063497816773886
-LanguageModel_OOV	-0.1146750000000015
-PhraseModel_0	-0.062922130123762257
-PhraseModel_1	0.0035552404454581212
-PhraseModel_2	0.039691524494244249
-PhraseModel_3	0.080265456972269417
-PhraseModel_4	-0.57787128729945014
-PhraseModel_5	0.017399999999999922
-PhraseModel_6	-0.17095000000000066
+WordPenalty 0.11138567724613679
+LanguageModel 0.95438136276453733
+LanguageModel_OOV -0.060799999999999937
+PhraseModel_0 -0.98112865741560529
+PhraseModel_1 -0.090531125075232435
+PhraseModel_2 0.79088062624556033
+PhraseModel_3 -0.57623134776057228
+PhraseModel_4 -1.4382448344095151
+PhraseModel_5 0.02435
+PhraseModel_6 -0.108125
+Glue 0.31832499999999997
+PassThrough -0.090950000000000003
diff --git a/training/dtrain/examples/parallelized/work/weights.3.2 b/training/dtrain/examples/parallelized/work/weights.3.2
index 3ff0411d..ccb591a2 100644
--- a/training/dtrain/examples/parallelized/work/weights.3.2
+++ b/training/dtrain/examples/parallelized/work/weights.3.2
@@ -1,12 +1,12 @@
-PassThrough	-0.32552500000000006
-Glue	-0.13247499999999815
-WordPenalty	0.053591939066858545
-LanguageModel	0.72104728811924446
-LanguageModel_OOV	-0.30827499999999869
-PhraseModel_0	-0.37052837676792744
-PhraseModel_1	0.17551097460105014
-PhraseModel_2	0.49999630285778179
-PhraseModel_3	-0.14590465814428336
-PhraseModel_4	-0.59563132644367889
-PhraseModel_5	0.034750000000000433
-PhraseModel_6	-0.11142500000000025
+WordPenalty 0.13650961302423945
+LanguageModel 0.58946464694775647
+LanguageModel_OOV -0.48362499999999997
+PhraseModel_0 -0.81261645844738917
+PhraseModel_1 0.44272714074140529
+PhraseModel_2 1.1732783465445731
+PhraseModel_3 -0.18260393204552733
+PhraseModel_4 -1.2213298752899167
+PhraseModel_5 0.02435
+PhraseModel_6 -0.188225
+Glue -0.12992500000000001
+PassThrough -0.51377500000000009