Merge pull request #17 from pks/master

dtrain

66 files changed, 764 insertions, 556 deletions

diff --git a/training/dtrain/README.md b/training/dtrain/README.md
index 7edabbf1..2ab2f232 100644
--- a/training/dtrain/README.md
+++ b/training/dtrain/README.md
@@ -13,36 +13,18 @@ Builds when building cdec, see ../BUILDING .
 To build only parts needed for dtrain do
 ```
   autoreconf -ifv
-  ./configure [--disable-gtest]
-  cd dtrain/; make
+  ./configure
+  cd training/dtrain/; make
 ```
 
 Running
 -------
-To run this on a dev set locally:
-```
-    #define DTRAIN_LOCAL
-```
-otherwise remove that line or undef, then recompile. You need a single
-grammar file or input annotated with per-sentence grammars (psg) as you
-would use with cdec. Additionally you need to give dtrain a file with
-references (--refs) when running locally.
-
-The input for use with hadoop streaming looks like this:
-```
-    <sid>\t<source>\t<ref>\t<grammar rules separated by \t>
-```
-To convert a psg to this format you need to replace all "\n"
-by "\t". Make sure there are no tabs in your data.
-
-For an example of local usage (with the 'distributed' format)
-the see test/example/ . This expects dtrain to be built without
-DTRAIN_LOCAL.
+See directories under test/ .
 
 Legal
 -----
-Copyright (c) 2012 by Patrick Simianer <p@simianer.de>
+Copyright (c) 2012-2013 by Patrick Simianer <p@simianer.de>
 
-See the file ../LICENSE.txt for the licensing terms that this software is
+See the file LICENSE.txt in the root folder for the licensing terms that this software is
 released under.
 
diff --git a/training/dtrain/dtrain.cc b/training/dtrain/dtrain.cc
index 18286668..149f87d4 100644
--- a/training/dtrain/dtrain.cc
+++ b/training/dtrain/dtrain.cc
@@ -6,15 +6,14 @@ dtrain_init(int argc, char** argv, po::variables_map* cfg)
 {
   po::options_description ini("Configuration File Options");
   ini.add_options()
-    ("input",             po::value<string>()->default_value("-"),                                                   "input file")
+    ("input",             po::value<string>()->default_value("-"),                                             "input file (src)")
+    ("refs,r",            po::value<string>(),                                                                       "references")
     ("output",            po::value<string>()->default_value("-"),                          "output weights file, '-' for STDOUT")
     ("input_weights",     po::value<string>(),                                "input weights file (e.g. from previous iteration)")
     ("decoder_config",    po::value<string>(),                                                      "configuration file for cdec")
     ("print_weights",     po::value<string>(),                                               "weights to print on each iteration")
     ("stop_after",        po::value<unsigned>()->default_value(0),                                 "stop after X input sentences")
-    ("tmp",               po::value<string>()->default_value("/tmp"),                                           "temp dir to use")
     ("keep",              po::value<bool>()->zero_tokens(),                               "keep weights files for each iteration")
-    ("hstreaming",        po::value<string>(),                                   "run in hadoop streaming mode, arg is a task id")
     ("epochs",            po::value<unsigned>()->default_value(10),                               "# of iterations T (per shard)")
     ("k",                 po::value<unsigned>()->default_value(100),                            "how many translations to sample")
     ("sample_from",       po::value<string>()->default_value("kbest"),     "where to sample translations from: 'kbest', 'forest'")
@@ -28,16 +27,13 @@ dtrain_init(int argc, char** argv, po::variables_map* cfg)
     ("gamma",             po::value<weight_t>()->default_value(0.),                            "gamma for SVM (0 for perceptron)")
     ("select_weights",    po::value<string>()->default_value("last"),     "output best, last, avg weights ('VOID' to throw away)")
     ("rescale",           po::value<bool>()->zero_tokens(),                              "rescale weight vector after each input")
-    ("l1_reg",            po::value<string>()->default_value("none"),      "apply l1 regularization as in 'Tsuroka et al' (2010)")
+    ("l1_reg",            po::value<string>()->default_value("none"), "apply l1 regularization as in 'Tsuroka et al' (2010) UNTESTED")
     ("l1_reg_strength",   po::value<weight_t>(),                                                     "l1 regularization strength")
     ("fselect",           po::value<weight_t>()->default_value(-1), "select top x percent (or by threshold) of features after each epoch NOT IMPLEMENTED") // TODO
     ("approx_bleu_d",     po::value<score_t>()->default_value(0.9),                                   "discount for approx. BLEU")
     ("scale_bleu_diff",   po::value<bool>()->zero_tokens(),                      "learning rate <- bleu diff of a misranked pair")
     ("loss_margin",       po::value<weight_t>()->default_value(0.),  "update if no error in pref pair but model scores this near")
     ("max_pairs",         po::value<unsigned>()->default_value(std::numeric_limits<unsigned>::max()), "max. # of pairs per Sent.")
-#ifdef DTRAIN_LOCAL
-    ("refs,r",            po::value<string>(),                                                         "references in local mode")
-#endif
     ("noup",              po::value<bool>()->zero_tokens(),                                               "do not update weights");
   po::options_description cl("Command Line Options");
   cl.add_options()
@@ -55,16 +51,6 @@ dtrain_init(int argc, char** argv, po::variables_map* cfg)
     cerr << cl << endl;
     return false;
   }
-  if (cfg->count("hstreaming") && (*cfg)["output"].as<string>() != "-") {
-    cerr << "When using 'hstreaming' the 'output' param should be '-'." << endl;
-    return false;
-  }
-#ifdef DTRAIN_LOCAL
-  if ((*cfg)["input"].as<string>() == "-") {
-    cerr << "Can't use stdin as input with this binary. Recompile without DTRAIN_LOCAL" << endl;
-    return false;
-  }
-#endif
   if ((*cfg)["sample_from"].as<string>() != "kbest"
        && (*cfg)["sample_from"].as<string>() != "forest") {
     cerr << "Wrong 'sample_from' param: '" << (*cfg)["sample_from"].as<string>() << "', use 'kbest' or 'forest'." << endl;
@@ -111,17 +97,8 @@ main(int argc, char** argv)
   if (cfg.count("verbose")) verbose = true;
   bool noup = false;
   if (cfg.count("noup")) noup = true;
-  bool hstreaming = false;
-  string task_id;
-  if (cfg.count("hstreaming")) {
-    hstreaming = true;
-    quiet = true;
-    task_id = cfg["hstreaming"].as<string>();
-    cerr.precision(17);
-  }
   bool rescale = false;
   if (cfg.count("rescale")) rescale = true;
-  HSReporter rep(task_id);
   bool keep = false;
   if (cfg.count("keep")) keep = true;
 
@@ -148,6 +125,7 @@ main(int argc, char** argv)
   if (cfg.count("print_weights"))
     boost::split(print_weights, cfg["print_weights"].as<string>(), boost::is_any_of(" "));
 
+
   // setup decoder
   register_feature_functions();
   SetSilent(true);
@@ -163,6 +141,8 @@ main(int argc, char** argv)
     scorer = dynamic_cast<BleuScorer*>(new BleuScorer);
   } else if (scorer_str == "stupid_bleu") {
     scorer = dynamic_cast<StupidBleuScorer*>(new StupidBleuScorer);
+  } else if (scorer_str == "fixed_stupid_bleu") {
+    scorer = dynamic_cast<FixedStupidBleuScorer*>(new FixedStupidBleuScorer);
   } else if (scorer_str == "smooth_bleu") {
     scorer = dynamic_cast<SmoothBleuScorer*>(new SmoothBleuScorer);
   } else if (scorer_str == "sum_bleu") {
@@ -201,6 +181,11 @@ main(int argc, char** argv)
   weight_t eta = cfg["learning_rate"].as<weight_t>();
   weight_t gamma = cfg["gamma"].as<weight_t>();
 
+  // faster perceptron: consider only misranked pairs, see
+  // DO NOT ENABLE  WITH SVM (gamma > 0) OR loss_margin!
+  bool faster_perceptron = false;
+  if (gamma==0 && loss_margin==0) faster_perceptron = true;
+
   // l1 regularization
   bool l1naive = false;
   bool l1clip = false;
@@ -222,16 +207,8 @@ main(int argc, char** argv)
   // buffer input for t > 0
   vector<string> src_str_buf;          // source strings (decoder takes only strings)
   vector<vector<WordID> > ref_ids_buf; // references as WordID vecs
-  // where temp files go
-  string tmp_path = cfg["tmp"].as<string>();
-#ifdef DTRAIN_LOCAL
   string refs_fn = cfg["refs"].as<string>();
   ReadFile refs(refs_fn);
-#else
-  string grammar_buf_fn = gettmpf(tmp_path, "dtrain-grammars");
-  ogzstream grammar_buf_out;
-  grammar_buf_out.open(grammar_buf_fn.c_str());
-#endif
 
   unsigned in_sz = std::numeric_limits<unsigned>::max(); // input index, input size
   vector<pair<score_t, score_t> > all_scores;
@@ -246,7 +223,7 @@ main(int argc, char** argv)
     cerr << setw(25) << "k " << k << endl;
     cerr << setw(25) << "N " << N << endl;
     cerr << setw(25) << "T " << T << endl;
-    cerr << setw(25) << "scorer '" << scorer_str << "'" << endl;
+    cerr << setw(26) << "scorer '" << scorer_str << "'" << endl;
     if (scorer_str == "approx_bleu")
       cerr << setw(25) << "approx. B discount " << approx_bleu_d << endl;
     cerr << setw(25) << "sample from " << "'" << sample_from << "'" << endl;
@@ -256,6 +233,7 @@ main(int argc, char** argv)
     else cerr << setw(25) << "learning rate " << "bleu diff" << endl;
     cerr << setw(25) << "gamma " << gamma << endl;
     cerr << setw(25) << "loss margin " << loss_margin << endl;
+    cerr << setw(25) << "faster perceptron " << faster_perceptron << endl;
     cerr << setw(25) << "pairs " << "'" << pair_sampling << "'" << endl;
     if (pair_sampling == "XYX")
       cerr << setw(25) << "hi lo " << hi_lo << endl;
@@ -268,9 +246,7 @@ main(int argc, char** argv)
     cerr << setw(25) << "max pairs " << max_pairs << endl;
     cerr << setw(25) << "cdec cfg " << "'" << cfg["decoder_config"].as<string>() << "'" << endl;
     cerr << setw(25) << "input " << "'" << input_fn << "'" << endl;
-#ifdef DTRAIN_LOCAL
     cerr << setw(25) << "refs " << "'" << refs_fn << "'" << endl;
-#endif
     cerr << setw(25) << "output " << "'" << output_fn << "'" << endl;
     if (cfg.count("input_weights"))
       cerr << setw(25) << "weights in " << "'" << cfg["input_weights"].as<string>() << "'" << endl;
@@ -283,14 +259,8 @@ main(int argc, char** argv)
   for (unsigned t = 0; t < T; t++) // T epochs
   {
 
-  if (hstreaming) cerr << "reporter:status:Iteration #" << t+1 << " of " << T << endl;
-
   time_t start, end;
   time(&start);
-#ifndef DTRAIN_LOCAL
-  igzstream grammar_buf_in;
-  if (t > 0) grammar_buf_in.open(grammar_buf_fn.c_str());
-#endif
   score_t score_sum = 0.;
   score_t model_sum(0);
   unsigned ii = 0, rank_errors = 0, margin_violations = 0, npairs = 0, f_count = 0, list_sz = 0;
@@ -338,52 +308,6 @@ main(int argc, char** argv)
 
     // getting input
     vector<WordID> ref_ids; // reference as vector<WordID>
-#ifndef DTRAIN_LOCAL
-    vector<string> in_split; // input: sid\tsrc\tref\tpsg
-    if (t == 0) {
-      // handling input
-      split_in(in, in_split);
-      if (hstreaming && ii == 0) cerr << "reporter:counter:" << task_id << ",First ID," << in_split[0] << endl;
-      // getting reference
-      vector<string> ref_tok;
-      boost::split(ref_tok, in_split[2], boost::is_any_of(" "));
-      register_and_convert(ref_tok, ref_ids);
-      ref_ids_buf.push_back(ref_ids);
-      // process and set grammar
-      bool broken_grammar = true; // ignore broken grammars
-      for (string::iterator it = in.begin(); it != in.end(); it++) {
-        if (!isspace(*it)) {
-          broken_grammar = false;
-          break;
-        }
-      }
-      if (broken_grammar) {
-        cerr << "Broken grammar for " << ii+1 << "! Ignoring this input." << endl;
-        continue;
-      }
-      boost::replace_all(in, "\t", "\n");
-      in += "\n";
-      grammar_buf_out << in << DTRAIN_GRAMMAR_DELIM << " " << in_split[0] << endl;
-      decoder.AddSupplementalGrammarFromString(in);
-      src_str_buf.push_back(in_split[1]);
-      // decode
-      observer->SetRef(ref_ids);
-      decoder.Decode(in_split[1], observer);
-    } else {
-      // get buffered grammar
-      string grammar_str;
-      while (true) {
-        string rule;
-        getline(grammar_buf_in, rule);
-        if (boost::starts_with(rule, DTRAIN_GRAMMAR_DELIM)) break;
-        grammar_str += rule + "\n";
-      }
-      decoder.AddSupplementalGrammarFromString(grammar_str);
-      // decode
-      observer->SetRef(ref_ids_buf[ii]);
-      decoder.Decode(src_str_buf[ii], observer);
-    }
-#else
     if (t == 0) {
       string r_;
       getline(*refs, r_);
@@ -400,7 +324,6 @@ main(int argc, char** argv)
       decoder.Decode(in, observer);
     else
       decoder.Decode(src_str_buf[ii], observer);
-#endif
 
     // get (scored) samples
     vector<ScoredHyp>* samples = observer->GetSamples();
@@ -430,25 +353,26 @@ main(int argc, char** argv)
       // get pairs
       vector<pair<ScoredHyp,ScoredHyp> > pairs;
       if (pair_sampling == "all")
-        all_pairs(samples, pairs, pair_threshold, max_pairs);
+        all_pairs(samples, pairs, pair_threshold, max_pairs, faster_perceptron);
       if (pair_sampling == "XYX")
-        partXYX(samples, pairs, pair_threshold, max_pairs, hi_lo);
+        partXYX(samples, pairs, pair_threshold, max_pairs, faster_perceptron, hi_lo);
       if (pair_sampling == "PRO")
         PROsampling(samples, pairs, pair_threshold, max_pairs);
       npairs += pairs.size();
 
       for (vector<pair<ScoredHyp,ScoredHyp> >::iterator it = pairs.begin();
            it != pairs.end(); it++) {
-#ifdef DTRAIN_FASTER_PERCEPTRON
-        bool rank_error = true; // pair sampling already did this for us
-        rank_errors++;
-        score_t margin = std::numeric_limits<float>::max();
-#else
-        bool rank_error = it->first.model <= it->second.model;
+        bool rank_error;
+        score_t margin;
+        if (faster_perceptron) { // we only have considering misranked pairs
+          rank_error = true; // pair sampling already did this for us
+          margin = std::numeric_limits<float>::max();
+        } else {
+          rank_error = it->first.model <= it->second.model;
+          margin = fabs(fabs(it->first.model) - fabs(it->second.model));
+          if (!rank_error && margin < loss_margin) margin_violations++;
+        }
         if (rank_error) rank_errors++;
-        score_t margin = fabs(fabs(it->first.model) - fabs(it->second.model));
-        if (!rank_error && margin < loss_margin) margin_violations++;
-#endif
         if (scale_bleu_diff) eta = it->first.score - it->second.score;
         if (rank_error || margin < loss_margin) {
           SparseVector<weight_t> diff_vec = it->first.f - it->second.f;
@@ -459,35 +383,40 @@ main(int argc, char** argv)
       }
 
       // l1 regularization
+      // please note that this penalizes _all_ weights
+      // (contrary to only the ones changed by the last update)
+      // after a _sentence_ (not after each example/pair)
       if (l1naive) {
-        for (unsigned d = 0; d < lambdas.size(); d++) {
-          weight_t v = lambdas.get(d);
-          lambdas.set_value(d, v - sign(v) * l1_reg);
+        FastSparseVector<weight_t>::iterator it = lambdas.begin();
+        for (; it != lambdas.end(); ++it) {
+          it->second -= sign(it->second) * l1_reg;
         }
       } else if (l1clip) {
-        for (unsigned d = 0; d < lambdas.size(); d++) {
-          if (lambdas.nonzero(d)) {
-            weight_t v = lambdas.get(d);
+        FastSparseVector<weight_t>::iterator it = lambdas.begin();
+        for (; it != lambdas.end(); ++it) {
+          if (it->second != 0) {
+            weight_t v = it->second;
             if (v > 0) {
-              lambdas.set_value(d, max(0., v - l1_reg));
+              it->second = max(0., v - l1_reg);
             } else {
-              lambdas.set_value(d, min(0., v + l1_reg));
+              it->second = min(0., v + l1_reg);
             }
           }
         }
       } else if (l1cumul) {
         weight_t acc_penalty = (ii+1) * l1_reg; // ii is the index of the current input
-        for (unsigned d = 0; d < lambdas.size(); d++) {
-          if (lambdas.nonzero(d)) {
-            weight_t v = lambdas.get(d);
-            weight_t penalty = 0;
+        FastSparseVector<weight_t>::iterator it = lambdas.begin();
+        for (; it != lambdas.end(); ++it) {
+          if (it->second != 0) {
+            weight_t v = it->second;
+            weight_t penalized = 0.;
             if (v > 0) {
-              penalty = max(0., v-(acc_penalty + cumulative_penalties.get(d)));
+              penalized = max(0., v-(acc_penalty + cumulative_penalties.get(it->first)));
             } else {
-              penalty = min(0., v+(acc_penalty - cumulative_penalties.get(d)));
+              penalized = min(0., v+(acc_penalty - cumulative_penalties.get(it->first)));
             }
-            lambdas.set_value(d, penalty);
-            cumulative_penalties.set_value(d, cumulative_penalties.get(d)+penalty);
+            it->second = penalized;
+            cumulative_penalties.set_value(it->first, cumulative_penalties.get(it->first)+penalized);
           }
         }
       }
@@ -498,11 +427,6 @@ main(int argc, char** argv)
 
     ++ii;
 
-    if (hstreaming) {
-      rep.update_counter("Seen #"+boost::lexical_cast<string>(t+1), 1u);
-      rep.update_counter("Seen", 1u);
-    }
-
   } // input loop
 
   if (average) w_average += lambdas;
@@ -511,21 +435,8 @@ main(int argc, char** argv)
 
   if (t == 0) {
     in_sz = ii; // remember size of input (# lines)
-    if (hstreaming) {
-      rep.update_counter("|Input|", ii);
-      rep.update_gcounter("|Input|", ii);
-      rep.update_gcounter("Shards", 1u);
-    }
   }
 
-#ifndef DTRAIN_LOCAL
-  if (t == 0) {
-    grammar_buf_out.close();
-  } else {
-    grammar_buf_in.close();
-  }
-#endif
-
   // print some stats
   score_t score_avg = score_sum/(score_t)in_sz;
   score_t model_avg = model_sum/(score_t)in_sz;
@@ -539,7 +450,7 @@ main(int argc, char** argv)
   }
 
   unsigned nonz = 0;
-  if (!quiet || hstreaming) nonz = (unsigned)lambdas.num_nonzero();
+  if (!quiet) nonz = (unsigned)lambdas.num_nonzero();
 
   if (!quiet) {
     cerr << _p5 << _p << "WEIGHTS" << endl;
@@ -552,28 +463,18 @@ main(int argc, char** argv)
     cerr << _np << " 1best avg model score: " << model_avg;
     cerr << _p << " (" << model_diff << ")" << endl;
     cerr << "           avg # pairs: ";
-    cerr << _np << npairs/(float)in_sz << endl;
+    cerr << _np << npairs/(float)in_sz;
+    if (faster_perceptron) cerr << " (meaningless)";
+    cerr << endl;
     cerr << "        avg # rank err: ";
     cerr << rank_errors/(float)in_sz << endl;
-#ifndef DTRAIN_FASTER_PERCEPTRON
     cerr << "     avg # margin viol: ";
     cerr << margin_violations/(float)in_sz << endl;
-#endif
     cerr << "    non0 feature count: " <<  nonz << endl;
     cerr << "           avg list sz: " << list_sz/(float)in_sz << endl;
     cerr << "           avg f count: " << f_count/(float)list_sz << endl;
   }
 
-  if (hstreaming) {
-    rep.update_counter("Score 1best avg #"+boost::lexical_cast<string>(t+1), (unsigned)(score_avg*DTRAIN_SCALE));
-    rep.update_counter("Model 1best avg #"+boost::lexical_cast<string>(t+1), (unsigned)(model_avg*DTRAIN_SCALE));
-    rep.update_counter("Pairs avg #"+boost::lexical_cast<string>(t+1), (unsigned)((npairs/(weight_t)in_sz)*DTRAIN_SCALE));
-    rep.update_counter("Rank errors avg #"+boost::lexical_cast<string>(t+1), (unsigned)((rank_errors/(weight_t)in_sz)*DTRAIN_SCALE));
-    rep.update_counter("Margin violations avg #"+boost::lexical_cast<string>(t+1), (unsigned)((margin_violations/(weight_t)in_sz)*DTRAIN_SCALE));
-    rep.update_counter("Non zero feature count #"+boost::lexical_cast<string>(t+1), nonz);
-    rep.update_gcounter("Non zero feature count #"+boost::lexical_cast<string>(t+1), nonz);
-  }
-
   pair<score_t,score_t> remember;
   remember.first = score_avg;
   remember.second = model_avg;
@@ -604,10 +505,6 @@ main(int argc, char** argv)
 
   if (average) w_average /= (weight_t)T;
 
-#ifndef DTRAIN_LOCAL
-  unlink(grammar_buf_fn.c_str());
-#endif
-
   if (!noup) {
     if (!quiet) cerr << endl << "Writing weights file to '" << output_fn << "' ..." << endl;
     if (select_weights == "last" || average) { // last, average
@@ -644,7 +541,6 @@ main(int argc, char** argv)
         }
       }
     }
-    if (output_fn == "-" && hstreaming) cout << "__SHARD_COUNT__\t1" << endl;
     if (!quiet) cerr << "done" << endl;
   }
 
diff --git a/training/dtrain/dtrain.h b/training/dtrain/dtrain.h
index 4b6f415c..eb0b9f17 100644
--- a/training/dtrain/dtrain.h
+++ b/training/dtrain/dtrain.h
@@ -1,14 +1,9 @@
 #ifndef _DTRAIN_H_
 #define _DTRAIN_H_
 
-#undef DTRAIN_FASTER_PERCEPTRON // only look at misranked pairs
-                                 // DO NOT USE WITH SVM!
-//#define DTRAIN_LOCAL
 #define DTRAIN_DOTS 10 // after how many inputs to display a '.'
-#define DTRAIN_GRAMMAR_DELIM "########EOS########"
 #define DTRAIN_SCALE 100000
 
-
 #include <iomanip>
 #include <climits>
 #include <string.h>
diff --git a/training/dtrain/examples/parallelized/README b/training/dtrain/examples/parallelized/README
new file mode 100644
index 00000000..89715105
--- /dev/null
+++ b/training/dtrain/examples/parallelized/README
@@ -0,0 +1,5 @@
+run for example
+  ../../parallelize.rb ./dtrain.ini 4 false 2 2 ./in ./refs
+
+final weights will be in the file work/weights.3
+
diff --git a/training/dtrain/test/parallelize/cdec.ini b/training/dtrain/examples/parallelized/cdec.ini
index 72e99dc5..e43ba1c4 100644
--- a/training/dtrain/test/parallelize/cdec.ini
+++ b/training/dtrain/examples/parallelized/cdec.ini
@@ -4,7 +4,7 @@ intersection_strategy=cube_pruning
 cubepruning_pop_limit=200
 scfg_max_span_limit=15
 feature_function=WordPenalty
-feature_function=KLanguageModel /stor/dat/wmt12/en/news_only/m/wmt12.news.en.3.kenv5
+feature_function=KLanguageModel ../example/nc-wmt11.en.srilm.gz
 #feature_function=ArityPenalty
 #feature_function=CMR2008ReorderingFeatures
 #feature_function=Dwarf
diff --git a/training/dtrain/test/parallelize/dtrain.ini b/training/dtrain/examples/parallelized/dtrain.ini
index 03f9d240..f19ef891 100644
--- a/training/dtrain/test/parallelize/dtrain.ini
+++ b/training/dtrain/examples/parallelized/dtrain.ini
@@ -2,7 +2,7 @@ k=100
 N=4
 learning_rate=0.0001
 gamma=0
-loss_margin=0
+loss_margin=1.0
 epochs=1
 scorer=stupid_bleu
 sample_from=kbest
@@ -11,5 +11,6 @@ pair_sampling=XYX
 hi_lo=0.1
 select_weights=last
 print_weights=Glue WordPenalty LanguageModel LanguageModel_OOV PhraseModel_0 PhraseModel_1 PhraseModel_2 PhraseModel_3 PhraseModel_4 PhraseModel_5 PhraseModel_6 PassThrough
-tmp=/tmp
+# newer version of the grammar extractor use different feature names: 
+#print_weights=Glue WordPenalty LanguageModel LanguageModel_OOV PhraseModel_0 PhraseModel_1 PhraseModel_2 PhraseModel_3 PhraseModel_4 PhraseModel_5 PhraseModel_6 PassThrough
 decoder_config=cdec.ini
diff --git a/training/dtrain/examples/parallelized/grammar/grammar.out.0.gz b/training/dtrain/examples/parallelized/grammar/grammar.out.0.gz
new file mode 100644
index 00000000..1e28a24b
--- /dev/null
+++ b/training/dtrain/examples/parallelized/grammar/grammar.out.0.gz
diff --git a/training/dtrain/examples/parallelized/grammar/grammar.out.1.gz b/training/dtrain/examples/parallelized/grammar/grammar.out.1.gz
new file mode 100644
index 00000000..372f5675
--- /dev/null
+++ b/training/dtrain/examples/parallelized/grammar/grammar.out.1.gz
diff --git a/training/dtrain/examples/parallelized/grammar/grammar.out.2.gz b/training/dtrain/examples/parallelized/grammar/grammar.out.2.gz
new file mode 100644
index 00000000..145d0dc0
--- /dev/null
+++ b/training/dtrain/examples/parallelized/grammar/grammar.out.2.gz
diff --git a/training/dtrain/examples/parallelized/grammar/grammar.out.3.gz b/training/dtrain/examples/parallelized/grammar/grammar.out.3.gz
new file mode 100644
index 00000000..105593ff
--- /dev/null
+++ b/training/dtrain/examples/parallelized/grammar/grammar.out.3.gz
diff --git a/training/dtrain/examples/parallelized/grammar/grammar.out.4.gz b/training/dtrain/examples/parallelized/grammar/grammar.out.4.gz
new file mode 100644
index 00000000..30781f48
--- /dev/null
+++ b/training/dtrain/examples/parallelized/grammar/grammar.out.4.gz
diff --git a/training/dtrain/examples/parallelized/grammar/grammar.out.5.gz b/training/dtrain/examples/parallelized/grammar/grammar.out.5.gz
new file mode 100644
index 00000000..834ee759
--- /dev/null
+++ b/training/dtrain/examples/parallelized/grammar/grammar.out.5.gz
diff --git a/training/dtrain/examples/parallelized/grammar/grammar.out.6.gz b/training/dtrain/examples/parallelized/grammar/grammar.out.6.gz
new file mode 100644
index 00000000..2e76f348
--- /dev/null
+++ b/training/dtrain/examples/parallelized/grammar/grammar.out.6.gz
diff --git a/training/dtrain/examples/parallelized/grammar/grammar.out.7.gz b/training/dtrain/examples/parallelized/grammar/grammar.out.7.gz
new file mode 100644
index 00000000..3741a887
--- /dev/null
+++ b/training/dtrain/examples/parallelized/grammar/grammar.out.7.gz
diff --git a/training/dtrain/examples/parallelized/grammar/grammar.out.8.gz b/training/dtrain/examples/parallelized/grammar/grammar.out.8.gz
new file mode 100644
index 00000000..ebf6bd0c
--- /dev/null
+++ b/training/dtrain/examples/parallelized/grammar/grammar.out.8.gz
diff --git a/training/dtrain/examples/parallelized/grammar/grammar.out.9.gz b/training/dtrain/examples/parallelized/grammar/grammar.out.9.gz
new file mode 100644
index 00000000..c1791059
--- /dev/null
+++ b/training/dtrain/examples/parallelized/grammar/grammar.out.9.gz
diff --git a/training/dtrain/examples/parallelized/in b/training/dtrain/examples/parallelized/in
new file mode 100644
index 00000000..51d01fe7
--- /dev/null
+++ b/training/dtrain/examples/parallelized/in
@@ -0,0 +1,10 @@
+<seg grammar="grammar/grammar.out.0.gz" id="0">europas nach rassen geteiltes haus</seg>
+<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>
+<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>
+<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>
+<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>
+<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>
+<seg grammar="grammar/grammar.out.6.gz" id="6">das aber wird es nicht , wie die geschichte des rassismus in amerika deutlich zeigt .</seg>
+<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>
+<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>
+<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>
diff --git a/training/dtrain/examples/parallelized/refs b/training/dtrain/examples/parallelized/refs
new file mode 100644
index 00000000..632e27b0
--- /dev/null
+++ b/training/dtrain/examples/parallelized/refs
@@ -0,0 +1,10 @@
+europe 's divided racial house
+a common feature of europe 's extreme right is its racism and use of the immigration issue as a political wedge .
+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 .
+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 .
+an aging population at home and ever more open borders imply increasing racial fragmentation in european countries .
+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 .
+it will not , as america 's racial history clearly shows .
+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 .
+the first step to address racial politics is to understand the origin and consequences of racial animosity , even if it means uncovering unpleasant truths .
+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/out.0.0 b/training/dtrain/examples/parallelized/work/out.0.0
new file mode 100644
index 00000000..7a00ed0f
--- /dev/null
+++ b/training/dtrain/examples/parallelized/work/out.0.0
@@ -0,0 +1,61 @@
+                cdec cfg 'cdec.ini'
+Loading the LM will be faster if you build a binary file.
+Reading ../example/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 3121929377
+
+dtrain
+Parameters:
+                       k 100
+                       N 4
+                       T 1
+                  scorer 'stupid_bleu'
+             sample from 'kbest'
+                  filter 'uniq'
+           learning rate 0.0001
+                   gamma 0
+             loss margin 1
+                   pairs 'XYX'
+                   hi lo 0.1
+          pair threshold 0
+          select weights 'last'
+                  l1 reg 0 'none'
+               max pairs 4294967295
+                cdec cfg 'cdec.ini'
+                   input 'work/shard.0.0.in'
+                    refs 'work/shard.0.0.refs'
+                  output 'work/weights.0.0'
+(a dot represents 10 inputs)
+Iteration #1 of 1.
+  5
+WEIGHTS
+              Glue = +0.2663
+       WordPenalty = -0.0079042
+     LanguageModel = +0.44782
+ LanguageModel_OOV = -0.0401
+     PhraseModel_0 = -0.193
+     PhraseModel_1 = +0.71321
+     PhraseModel_2 = +0.85196
+     PhraseModel_3 = -0.43986
+     PhraseModel_4 = -0.44803
+     PhraseModel_5 = -0.0538
+     PhraseModel_6 = -0.1788
+       PassThrough = -0.1477
+        ---
+       1best avg score: 0.17521 (+0.17521)
+ 1best avg model score: 21.556 (+21.556)
+           avg # pairs: 1671.2
+        avg # rank err: 1118.6
+     avg # margin viol: 552.6
+    non0 feature count: 12
+           avg list sz: 100
+           avg f count: 11.32
+(time 0.37 min, 4.4 s/S)
+
+Writing weights file to 'work/weights.0.0' ...
+done
+
+---
+Best iteration: 1 [SCORE 'stupid_bleu'=0.17521].
+This took 0.36667 min.
diff --git a/training/dtrain/examples/parallelized/work/out.0.1 b/training/dtrain/examples/parallelized/work/out.0.1
new file mode 100644
index 00000000..e2bd6649
--- /dev/null
+++ b/training/dtrain/examples/parallelized/work/out.0.1
@@ -0,0 +1,62 @@
+                cdec cfg 'cdec.ini'
+Loading the LM will be faster if you build a binary file.
+Reading ../example/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 2767202922
+
+dtrain
+Parameters:
+                       k 100
+                       N 4
+                       T 1
+                  scorer 'stupid_bleu'
+             sample from 'kbest'
+                  filter 'uniq'
+           learning rate 0.0001
+                   gamma 0
+             loss margin 1
+                   pairs 'XYX'
+                   hi lo 0.1
+          pair threshold 0
+          select weights 'last'
+                  l1 reg 0 'none'
+               max pairs 4294967295
+                cdec cfg 'cdec.ini'
+                   input 'work/shard.0.0.in'
+                    refs 'work/shard.0.0.refs'
+                  output 'work/weights.0.1'
+              weights in 'work/weights.0'
+(a dot represents 10 inputs)
+Iteration #1 of 1.
+  5
+WEIGHTS
+              Glue = -0.2699
+       WordPenalty = +0.080605
+     LanguageModel = -0.026572
+ LanguageModel_OOV = -0.30025
+     PhraseModel_0 = -0.32076
+     PhraseModel_1 = +0.67451
+     PhraseModel_2 = +0.92
+     PhraseModel_3 = -0.36402
+     PhraseModel_4 = -0.592
+     PhraseModel_5 = -0.0269
+     PhraseModel_6 = -0.28755
+       PassThrough = -0.33285
+        ---
+       1best avg score: 0.26638 (+0.26638)
+ 1best avg model score: 53.197 (+53.197)
+           avg # pairs: 2028.6
+        avg # rank err: 998.2
+     avg # margin viol: 918.8
+    non0 feature count: 12
+           avg list sz: 100
+           avg f count: 10.496
+(time 0.32 min, 3.8 s/S)
+
+Writing weights file to 'work/weights.0.1' ...
+done
+
+---
+Best iteration: 1 [SCORE 'stupid_bleu'=0.26638].
+This took 0.31667 min.
diff --git a/training/dtrain/examples/parallelized/work/out.1.0 b/training/dtrain/examples/parallelized/work/out.1.0
new file mode 100644
index 00000000..6e790e38
--- /dev/null
+++ b/training/dtrain/examples/parallelized/work/out.1.0
@@ -0,0 +1,61 @@
+                cdec cfg 'cdec.ini'
+Loading the LM will be faster if you build a binary file.
+Reading ../example/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 1432415010
+
+dtrain
+Parameters:
+                       k 100
+                       N 4
+                       T 1
+                  scorer 'stupid_bleu'
+             sample from 'kbest'
+                  filter 'uniq'
+           learning rate 0.0001
+                   gamma 0
+             loss margin 1
+                   pairs 'XYX'
+                   hi lo 0.1
+          pair threshold 0
+          select weights 'last'
+                  l1 reg 0 'none'
+               max pairs 4294967295
+                cdec cfg 'cdec.ini'
+                   input 'work/shard.1.0.in'
+                    refs 'work/shard.1.0.refs'
+                  output 'work/weights.1.0'
+(a dot represents 10 inputs)
+Iteration #1 of 1.
+  5
+WEIGHTS
+              Glue = -0.3815
+       WordPenalty = +0.20064
+     LanguageModel = +0.95304
+ LanguageModel_OOV = -0.264
+     PhraseModel_0 = -0.22362
+     PhraseModel_1 = +0.12254
+     PhraseModel_2 = +0.26328
+     PhraseModel_3 = +0.38018
+     PhraseModel_4 = -0.48654
+     PhraseModel_5 = +0
+     PhraseModel_6 = -0.3645
+       PassThrough = -0.2216
+        ---
+       1best avg score: 0.10863 (+0.10863)
+ 1best avg model score: -4.9841 (-4.9841)
+           avg # pairs: 1345.4
+        avg # rank err: 822.4
+     avg # margin viol: 501
+    non0 feature count: 11
+           avg list sz: 100
+           avg f count: 11.814
+(time 0.45 min, 5.4 s/S)
+
+Writing weights file to 'work/weights.1.0' ...
+done
+
+---
+Best iteration: 1 [SCORE 'stupid_bleu'=0.10863].
+This took 0.45 min.
diff --git a/training/dtrain/examples/parallelized/work/out.1.1 b/training/dtrain/examples/parallelized/work/out.1.1
new file mode 100644
index 00000000..0b984761
--- /dev/null
+++ b/training/dtrain/examples/parallelized/work/out.1.1
@@ -0,0 +1,62 @@
+                cdec cfg 'cdec.ini'
+Loading the LM will be faster if you build a binary file.
+Reading ../example/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 1771918374
+
+dtrain
+Parameters:
+                       k 100
+                       N 4
+                       T 1
+                  scorer 'stupid_bleu'
+             sample from 'kbest'
+                  filter 'uniq'
+           learning rate 0.0001
+                   gamma 0
+             loss margin 1
+                   pairs 'XYX'
+                   hi lo 0.1
+          pair threshold 0
+          select weights 'last'
+                  l1 reg 0 'none'
+               max pairs 4294967295
+                cdec cfg 'cdec.ini'
+                   input 'work/shard.1.0.in'
+                    refs 'work/shard.1.0.refs'
+                  output 'work/weights.1.1'
+              weights in 'work/weights.0'
+(a dot represents 10 inputs)
+Iteration #1 of 1.
+  5
+WEIGHTS
+              Glue = -0.3178
+       WordPenalty = +0.11092
+     LanguageModel = +0.17269
+ LanguageModel_OOV = -0.13485
+     PhraseModel_0 = -0.45371
+     PhraseModel_1 = +0.38789
+     PhraseModel_2 = +0.75311
+     PhraseModel_3 = -0.38163
+     PhraseModel_4 = -0.58817
+     PhraseModel_5 = -0.0269
+     PhraseModel_6 = -0.27315
+       PassThrough = -0.16745
+        ---
+       1best avg score: 0.13169 (+0.13169)
+ 1best avg model score: 24.226 (+24.226)
+           avg # pairs: 1951.2
+        avg # rank err: 985.4
+     avg # margin viol: 951
+    non0 feature count: 12
+           avg list sz: 100
+           avg f count: 11.224
+(time 0.42 min, 5 s/S)
+
+Writing weights file to 'work/weights.1.1' ...
+done
+
+---
+Best iteration: 1 [SCORE 'stupid_bleu'=0.13169].
+This took 0.41667 min.
diff --git a/training/dtrain/examples/parallelized/work/shard.0.0.in b/training/dtrain/examples/parallelized/work/shard.0.0.in
new file mode 100644
index 00000000..92f9c78e
--- /dev/null
+++ b/training/dtrain/examples/parallelized/work/shard.0.0.in
@@ -0,0 +1,5 @@
+<seg grammar="grammar/grammar.out.0.gz" id="0">europas nach rassen geteiltes haus</seg>
+<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>
+<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>
+<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>
+<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>
diff --git a/training/dtrain/examples/parallelized/work/shard.0.0.refs b/training/dtrain/examples/parallelized/work/shard.0.0.refs
new file mode 100644
index 00000000..bef68fee
--- /dev/null
+++ b/training/dtrain/examples/parallelized/work/shard.0.0.refs
@@ -0,0 +1,5 @@
+europe 's divided racial house
+a common feature of europe 's extreme right is its racism and use of the immigration issue as a political wedge .
+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 .
+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 .
+an aging population at home and ever more open borders imply increasing racial fragmentation in european countries .
diff --git a/training/dtrain/examples/parallelized/work/shard.1.0.in b/training/dtrain/examples/parallelized/work/shard.1.0.in
new file mode 100644
index 00000000..b7695ce7
--- /dev/null
+++ b/training/dtrain/examples/parallelized/work/shard.1.0.in
@@ -0,0 +1,5 @@
+<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>
+<seg grammar="grammar/grammar.out.6.gz" id="6">das aber wird es nicht , wie die geschichte des rassismus in amerika deutlich zeigt .</seg>
+<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>
+<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>
+<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>
diff --git a/training/dtrain/examples/parallelized/work/shard.1.0.refs b/training/dtrain/examples/parallelized/work/shard.1.0.refs
new file mode 100644
index 00000000..6076f6d5
--- /dev/null
+++ b/training/dtrain/examples/parallelized/work/shard.1.0.refs
@@ -0,0 +1,5 @@
+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 .
+it will not , as america 's racial history clearly shows .
+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 .
+the first step to address racial politics is to understand the origin and consequences of racial animosity , even if it means uncovering unpleasant truths .
+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
new file mode 100644
index 00000000..ddd595a8
--- /dev/null
+++ b/training/dtrain/examples/parallelized/work/weights.0
@@ -0,0 +1,12 @@
+LanguageModel	0.7004298992212881
+PhraseModel_2	0.5576194336478857
+PhraseModel_1	0.41787318415343155
+PhraseModel_4	-0.46728502545635164
+PhraseModel_3	-0.029839521598455515
+Glue	-0.05760000000000068
+PhraseModel_6	-0.2716499999999978
+PhraseModel_0	-0.20831031065605327
+LanguageModel_OOV	-0.15205000000000077
+PassThrough	-0.1846500000000006
+WordPenalty	0.09636994553433414
+PhraseModel_5	-0.026900000000000257
diff --git a/training/dtrain/examples/parallelized/work/weights.0.0 b/training/dtrain/examples/parallelized/work/weights.0.0
new file mode 100644
index 00000000..c9370b18
--- /dev/null
+++ b/training/dtrain/examples/parallelized/work/weights.0.0
@@ -0,0 +1,12 @@
+WordPenalty	-0.0079041595706392243
+LanguageModel	0.44781580828279532
+LanguageModel_OOV	-0.04010000000000042
+Glue	0.26629999999999948
+PhraseModel_0	-0.19299677809125185
+PhraseModel_1	0.71321026861732773
+PhraseModel_2	0.85195540993310537
+PhraseModel_3	-0.43986310822842656
+PhraseModel_4	-0.44802855630415955
+PhraseModel_5	-0.053800000000000514
+PhraseModel_6	-0.17879999999999835
+PassThrough	-0.14770000000000036
diff --git a/training/dtrain/examples/parallelized/work/weights.0.1 b/training/dtrain/examples/parallelized/work/weights.0.1
new file mode 100644
index 00000000..8fad3de8
--- /dev/null
+++ b/training/dtrain/examples/parallelized/work/weights.0.1
@@ -0,0 +1,12 @@
+WordPenalty	0.080605055841244472
+LanguageModel	-0.026571720531022844
+LanguageModel_OOV	-0.30024999999999141
+Glue	-0.26989999999999842
+PhraseModel_2	0.92000295209089566
+PhraseModel_1	0.67450748692470841
+PhraseModel_4	-0.5920000014976784
+PhraseModel_3	-0.36402437203127397
+PhraseModel_6	-0.28754999999999603
+PhraseModel_0	-0.32076244202907672
+PassThrough	-0.33284999999999004
+PhraseModel_5	-0.026900000000000257
diff --git a/training/dtrain/examples/parallelized/work/weights.1 b/training/dtrain/examples/parallelized/work/weights.1
new file mode 100644
index 00000000..03058a16
--- /dev/null
+++ b/training/dtrain/examples/parallelized/work/weights.1
@@ -0,0 +1,12 @@
+PhraseModel_2	0.8365578543552836
+PhraseModel_4	-0.5900840266009169
+PhraseModel_1	0.5312000609786991
+PhraseModel_0	-0.3872342271319619
+PhraseModel_3	-0.3728279676912084
+Glue	-0.2938500000000036
+PhraseModel_6	-0.2803499999999967
+PassThrough	-0.25014999999999626
+LanguageModel_OOV	-0.21754999999999702
+LanguageModel	0.07306061161169894
+WordPenalty	0.09576193325966899
+PhraseModel_5	-0.026900000000000257
diff --git a/training/dtrain/examples/parallelized/work/weights.1.0 b/training/dtrain/examples/parallelized/work/weights.1.0
new file mode 100644
index 00000000..6a6a65c1
--- /dev/null
+++ b/training/dtrain/examples/parallelized/work/weights.1.0
@@ -0,0 +1,11 @@
+WordPenalty	0.20064405063930751
+LanguageModel	0.9530439901597807
+LanguageModel_OOV	-0.26400000000000112
+Glue	-0.38150000000000084
+PhraseModel_0	-0.22362384322085468
+PhraseModel_1	0.12253609968953538
+PhraseModel_2	0.26328345736266612
+PhraseModel_3	0.38018406503151553
+PhraseModel_4	-0.48654149460854373
+PhraseModel_6	-0.36449999999999722
+PassThrough	-0.22160000000000085
diff --git a/training/dtrain/examples/parallelized/work/weights.1.1 b/training/dtrain/examples/parallelized/work/weights.1.1
new file mode 100644
index 00000000..f56ea4a2
--- /dev/null
+++ b/training/dtrain/examples/parallelized/work/weights.1.1
@@ -0,0 +1,12 @@
+WordPenalty	0.1109188106780935
+LanguageModel	0.17269294375442074
+LanguageModel_OOV	-0.13485000000000266
+Glue	-0.3178000000000088
+PhraseModel_2	0.75311275661967159
+PhraseModel_1	0.38789263503268989
+PhraseModel_4	-0.58816805170415531
+PhraseModel_3	-0.38163156335114284
+PhraseModel_6	-0.27314999999999739
+PhraseModel_0	-0.45370601223484697
+PassThrough	-0.16745000000000249
+PhraseModel_5	-0.026900000000000257
diff --git a/training/dtrain/examples/standard/README b/training/dtrain/examples/standard/README
new file mode 100644
index 00000000..ce37d31a
--- /dev/null
+++ b/training/dtrain/examples/standard/README
@@ -0,0 +1,2 @@
+Call `dtrain` from this folder with ../../dtrain -c dtrain.ini .
+
diff --git a/training/dtrain/test/example/cdec.ini b/training/dtrain/examples/standard/cdec.ini
index d5955f0e..e1edc68d 100644
--- a/training/dtrain/test/example/cdec.ini
+++ b/training/dtrain/examples/standard/cdec.ini
@@ -2,9 +2,10 @@ formalism=scfg
 add_pass_through_rules=true
 scfg_max_span_limit=15
 intersection_strategy=cube_pruning
-cubepruning_pop_limit=30
+cubepruning_pop_limit=200
+grammar=nc-wmt11.grammar.gz
 feature_function=WordPenalty
-feature_function=KLanguageModel test/example/nc-wmt11.en.srilm.gz
+feature_function=KLanguageModel ./nc-wmt11.en.srilm.gz
 # all currently working feature functions for translation:
 # (with those features active that were used in the ACL paper)
 #feature_function=ArityPenalty
diff --git a/training/dtrain/test/example/dtrain.ini b/training/dtrain/examples/standard/dtrain.ini
index 72d50ca1..e1072d30 100644
--- a/training/dtrain/test/example/dtrain.ini
+++ b/training/dtrain/examples/standard/dtrain.ini
@@ -1,10 +1,12 @@
-input=test/example/nc-wmt11.1k.gz    # use '-' for STDIN
-output=-                             # a weights file (add .gz for gzip compression) or STDOUT '-'
-select_weights=VOID                  # don't output weights
-decoder_config=test/example/cdec.ini # config for cdec
+input=./nc-wmt11.de.gz
+refs=./nc-wmt11.en.gz
+output=-                  # a weights file (add .gz for gzip compression) or STDOUT '-'
+select_weights=VOID       # output average (over epochs) weight vector
+decoder_config=./cdec.ini # config for cdec
 # weights for these features will be printed on each iteration
 print_weights=Glue WordPenalty LanguageModel LanguageModel_OOV PhraseModel_0 PhraseModel_1 PhraseModel_2 PhraseModel_3 PhraseModel_4 PhraseModel_5 PhraseModel_6 PassThrough
-tmp=/tmp
+# newer version of the grammar extractor use different feature names: 
+#print_weights= EgivenFCoherent SampleCountF CountEF MaxLexFgivenE MaxLexEgivenF IsSingletonF IsSingletonFE Glue WordPenalty PassThrough LanguageModel LanguageModel_OOV
 stop_after=10 # stop epoch after 10 inputs
 
 # interesting stuff
@@ -16,7 +18,7 @@ learning_rate=1.0       # learning rate, don't care if gamma=0 (perceptron)
 gamma=0                 # use SVM reg
 sample_from=kbest       # use kbest lists (as opposed to forest)
 filter=uniq             # only unique entries in kbest (surface form)
-pair_sampling=XYX
+pair_sampling=XYX       #
 hi_lo=0.1               # 10 vs 80 vs 10 and 80 vs 10 here
-pair_threshold=0        # minimum distance in BLEU (this will still only use pairs with diff > 0)
-loss_margin=0
+pair_threshold=0        # minimum distance in BLEU (here: > 0)
+loss_margin=0           # update if correctly ranked, but within this margin
diff --git a/training/dtrain/examples/standard/expected-output b/training/dtrain/examples/standard/expected-output
new file mode 100644
index 00000000..7cd09dbf
--- /dev/null
+++ b/training/dtrain/examples/standard/expected-output
@@ -0,0 +1,91 @@
+                cdec cfg './cdec.ini'
+Loading the LM will be faster if you build a binary file.
+Reading ./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
+****************************************************************************************************
+  Example feature: Shape_S00000_T00000
+Seeding random number sequence to 2679584485
+
+dtrain
+Parameters:
+                       k 100
+                       N 4
+                       T 2
+                  scorer 'stupid_bleu'
+             sample from 'kbest'
+                  filter 'uniq'
+           learning rate 1
+                   gamma 0
+             loss margin 0
+       faster perceptron 1
+                   pairs 'XYX'
+                   hi lo 0.1
+          pair threshold 0
+          select weights 'VOID'
+                  l1 reg 0 'none'
+               max pairs 4294967295
+                cdec cfg './cdec.ini'
+                   input './nc-wmt11.de.gz'
+                    refs './nc-wmt11.en.gz'
+                  output '-'
+              stop_after 10
+(a dot represents 10 inputs)
+Iteration #1 of 2.
+ . 10
+Stopping after 10 input sentences.
+WEIGHTS
+              Glue = -576
+       WordPenalty = +417.79
+     LanguageModel = +5117.5
+ LanguageModel_OOV = -1307
+     PhraseModel_0 = -1612
+     PhraseModel_1 = -2159.6
+     PhraseModel_2 = -677.36
+     PhraseModel_3 = +2663.8
+     PhraseModel_4 = -1025.9
+     PhraseModel_5 = -8
+     PhraseModel_6 = +70
+       PassThrough = -1455
+        ---
+       1best avg score: 0.27697 (+0.27697)
+ 1best avg model score: -47918 (-47918)
+           avg # pairs: 581.9 (meaningless)
+        avg # rank err: 581.9
+     avg # margin viol: 0
+    non0 feature count: 703
+           avg list sz: 90.9
+           avg f count: 100.09
+(time 0.25 min, 1.5 s/S)
+
+Iteration #2 of 2.
+ . 10
+WEIGHTS
+              Glue = -622
+       WordPenalty = +898.56
+     LanguageModel = +8066.2
+ LanguageModel_OOV = -2590
+     PhraseModel_0 = -4335.8
+     PhraseModel_1 = -5864.4
+     PhraseModel_2 = -1729.8
+     PhraseModel_3 = +2831.9
+     PhraseModel_4 = -5384.8
+     PhraseModel_5 = +1449
+     PhraseModel_6 = +480
+       PassThrough = -2578
+        ---
+       1best avg score: 0.37119 (+0.094226)
+ 1best avg model score: -1.3174e+05 (-83822)
+           avg # pairs: 584.1 (meaningless)
+        avg # rank err: 584.1
+     avg # margin viol: 0
+    non0 feature count: 1115
+           avg list sz: 91.3
+           avg f count: 90.755
+(time 0.3 min, 1.8 s/S)
+
+Writing weights file to '-' ...
+done
+
+---
+Best iteration: 2 [SCORE 'stupid_bleu'=0.37119].
+This took 0.55 min.
diff --git a/training/dtrain/examples/standard/nc-wmt11.de.gz b/training/dtrain/examples/standard/nc-wmt11.de.gz
new file mode 100644
index 00000000..0741fd92
--- /dev/null
+++ b/training/dtrain/examples/standard/nc-wmt11.de.gz
diff --git a/training/dtrain/examples/standard/nc-wmt11.en.gz b/training/dtrain/examples/standard/nc-wmt11.en.gz
new file mode 100644
index 00000000..1c0bd401
--- /dev/null
+++ b/training/dtrain/examples/standard/nc-wmt11.en.gz
diff --git a/training/dtrain/examples/standard/nc-wmt11.en.srilm.gz b/training/dtrain/examples/standard/nc-wmt11.en.srilm.gz
new file mode 100644
index 00000000..7ce81057
--- /dev/null
+++ b/training/dtrain/examples/standard/nc-wmt11.en.srilm.gz
diff --git a/training/dtrain/examples/standard/nc-wmt11.grammar.gz b/training/dtrain/examples/standard/nc-wmt11.grammar.gz
new file mode 100644
index 00000000..ce4024a1
--- /dev/null
+++ b/training/dtrain/examples/standard/nc-wmt11.grammar.gz
diff --git a/training/dtrain/test/toy/cdec.ini b/training/dtrain/examples/toy/cdec.ini
index 98b02d44..b14f4819 100644
--- a/training/dtrain/test/toy/cdec.ini
+++ b/training/dtrain/examples/toy/cdec.ini
@@ -1,2 +1,3 @@
 formalism=scfg
 add_pass_through_rules=true
+grammar=grammar.gz
diff --git a/training/dtrain/test/toy/dtrain.ini b/training/dtrain/examples/toy/dtrain.ini
index a091732f..cd715f26 100644
--- a/training/dtrain/test/toy/dtrain.ini
+++ b/training/dtrain/examples/toy/dtrain.ini
@@ -1,5 +1,6 @@
-decoder_config=test/toy/cdec.ini
-input=test/toy/input
+decoder_config=cdec.ini
+input=src
+refs=tgt
 output=-
 print_weights=logp shell_rule house_rule small_rule little_rule PassThrough
 k=4
diff --git a/training/dtrain/examples/toy/expected-output b/training/dtrain/examples/toy/expected-output
new file mode 100644
index 00000000..1da2aadd
--- /dev/null
+++ b/training/dtrain/examples/toy/expected-output
@@ -0,0 +1,77 @@
+Warning: hi_lo only works with pair_sampling XYX.
+                cdec cfg 'cdec.ini'
+Seeding random number sequence to 1664825829
+
+dtrain
+Parameters:
+                       k 4
+                       N 4
+                       T 2
+                  scorer 'bleu'
+             sample from 'kbest'
+                  filter 'uniq'
+           learning rate 1
+                   gamma 0
+             loss margin 0
+                   pairs 'all'
+          pair threshold 0
+          select weights 'last'
+                  l1 reg 0 'none'
+               max pairs 4294967295
+                cdec cfg 'cdec.ini'
+                   input 'src'
+                    refs 'tgt'
+                  output '-'
+(a dot represents 10 inputs)
+Iteration #1 of 2.
+  2
+WEIGHTS
+              logp = +0
+        shell_rule = -1
+        house_rule = +2
+        small_rule = -2
+       little_rule = +3
+       PassThrough = -5
+        ---
+       1best avg score: 0.5 (+0.5)
+ 1best avg model score: 2.5 (+2.5)
+           avg # pairs: 4
+        avg # rank err: 1.5
+     avg # margin viol: 0
+    non0 feature count: 6
+           avg list sz: 4
+           avg f count: 2.875
+(time 0 min, 0 s/S)
+
+Iteration #2 of 2.
+  2
+WEIGHTS
+              logp = +0
+        shell_rule = -1
+        house_rule = +2
+        small_rule = -2
+       little_rule = +3
+       PassThrough = -5
+        ---
+       1best avg score: 1 (+0.5)
+ 1best avg model score: 5 (+2.5)
+           avg # pairs: 5
+        avg # rank err: 0
+     avg # margin viol: 0
+    non0 feature count: 6
+           avg list sz: 4
+           avg f count: 3
+(time 0 min, 0 s/S)
+
+Writing weights file to '-' ...
+house_rule	2
+little_rule	3
+Glue	-4
+PassThrough	-5
+small_rule	-2
+shell_rule	-1
+done
+
+---
+Best iteration: 2 [SCORE 'bleu'=1].
+This took 0 min.
diff --git a/training/dtrain/examples/toy/grammar.gz b/training/dtrain/examples/toy/grammar.gz
new file mode 100644
index 00000000..8eb0d29e
--- /dev/null
+++ b/training/dtrain/examples/toy/grammar.gz
diff --git a/training/dtrain/examples/toy/src b/training/dtrain/examples/toy/src
new file mode 100644
index 00000000..87e39ef2
--- /dev/null
+++ b/training/dtrain/examples/toy/src
@@ -0,0 +1,2 @@
+ich sah ein kleines haus
+ich fand ein kleines haus
diff --git a/training/dtrain/examples/toy/tgt b/training/dtrain/examples/toy/tgt
new file mode 100644
index 00000000..174926b3
--- /dev/null
+++ b/training/dtrain/examples/toy/tgt
@@ -0,0 +1,2 @@
+i saw a little house
+i found a little house
diff --git a/training/dtrain/hstreaming/avg.rb b/training/dtrain/hstreaming/avg.rb
deleted file mode 100755
index 2599c732..00000000
--- a/training/dtrain/hstreaming/avg.rb
+++ /dev/null
@@ -1,32 +0,0 @@
-#!/usr/bin/env ruby
-# first arg may be an int of custom shard count
-
-shard_count_key = "__SHARD_COUNT__"
-
-STDIN.set_encoding 'utf-8'
-STDOUT.set_encoding 'utf-8'
-
-w = {}
-c = {}
-w.default = 0
-c.default = 0
-while line = STDIN.gets
-  key, val = line.split /\s/
-  w[key] += val.to_f
-  c[key] += 1
-end
-
-if ARGV.size == 0
-  shard_count = w["__SHARD_COUNT__"]
-else
-  shard_count = ARGV[0].to_f
-end
-w.each_key { |k|
-  if k == shard_count_key
-    next
-  else
-    puts "#{k}\t#{w[k]/shard_count}"
-    #puts "# #{c[k]}"
-  end
-}
-
diff --git a/training/dtrain/hstreaming/cdec.ini b/training/dtrain/hstreaming/cdec.ini
deleted file mode 100644
index d4f5cecd..00000000
--- a/training/dtrain/hstreaming/cdec.ini
+++ /dev/null
@@ -1,22 +0,0 @@
-formalism=scfg
-add_pass_through_rules=true
-scfg_max_span_limit=15
-intersection_strategy=cube_pruning
-cubepruning_pop_limit=30
-feature_function=WordPenalty
-feature_function=KLanguageModel nc-wmt11.en.srilm.gz
-#feature_function=ArityPenalty
-#feature_function=CMR2008ReorderingFeatures
-#feature_function=Dwarf
-#feature_function=InputIndicator
-#feature_function=LexNullJump
-#feature_function=NewJump
-#feature_function=NgramFeatures
-#feature_function=NonLatinCount
-#feature_function=OutputIndicator
-#feature_function=RuleIdentityFeatures
-#feature_function=RuleNgramFeatures
-#feature_function=RuleShape
-#feature_function=SourceSpanSizeFeatures
-#feature_function=SourceWordPenalty
-#feature_function=SpanFeatures
diff --git a/training/dtrain/hstreaming/dtrain.ini b/training/dtrain/hstreaming/dtrain.ini
deleted file mode 100644
index a2c219a1..00000000
--- a/training/dtrain/hstreaming/dtrain.ini
+++ /dev/null
@@ -1,15 +0,0 @@
-input=-
-output=-
-decoder_config=cdec.ini
-tmp=/var/hadoop/mapred/local/
-epochs=1
-k=100
-N=4
-learning_rate=0.0001
-gamma=0
-scorer=stupid_bleu
-sample_from=kbest
-filter=uniq
-pair_sampling=XYX
-pair_threshold=0
-select_weights=last
diff --git a/training/dtrain/hstreaming/dtrain.sh b/training/dtrain/hstreaming/dtrain.sh
deleted file mode 100755
index 877ff94c..00000000
--- a/training/dtrain/hstreaming/dtrain.sh
+++ /dev/null
@@ -1,9 +0,0 @@
-#!/bin/bash
-# script to run dtrain with a task id
-
-pushd . &>/dev/null
-cd ..
-ID=$(basename $(pwd)) # attempt_...
-popd &>/dev/null
-./dtrain -c dtrain.ini --hstreaming $ID
-
diff --git a/training/dtrain/hstreaming/hadoop-streaming-job.sh b/training/dtrain/hstreaming/hadoop-streaming-job.sh
deleted file mode 100755
index 92419956..00000000
--- a/training/dtrain/hstreaming/hadoop-streaming-job.sh
+++ /dev/null
@@ -1,30 +0,0 @@
-#!/bin/sh
-
-EXP=a_simple_test
-
-# change these vars to fit your hadoop installation
-HADOOP_HOME=/usr/lib/hadoop-0.20
-JAR=contrib/streaming/hadoop-streaming-0.20.2-cdh3u1.jar
-HSTREAMING="$HADOOP_HOME/bin/hadoop jar $HADOOP_HOME/$JAR"
-
- IN=input_on_hdfs
-OUT=output_weights_on_hdfs
-
-# you can -reducer to NONE if you want to
-# do feature selection/averaging locally (e.g. to
-# keep weights of all epochs)
-$HSTREAMING \
-    -mapper "dtrain.sh" \
-    -reducer "ruby lplp.rb l2 select_k 100000" \
-    -input $IN \
-    -output $OUT \
-    -file dtrain.sh \
-    -file lplp.rb \
-    -file ../dtrain \
-    -file dtrain.ini \
-    -file cdec.ini \
-    -file ../test/example/nc-wmt11.en.srilm.gz \
-    -jobconf mapred.reduce.tasks=30 \
-    -jobconf mapred.max.map.failures.percent=0 \
-    -jobconf mapred.job.name="dtrain $EXP"
-
diff --git a/training/dtrain/hstreaming/red-test b/training/dtrain/hstreaming/red-test
deleted file mode 100644
index 2623d697..00000000
--- a/training/dtrain/hstreaming/red-test
+++ /dev/null
@@ -1,9 +0,0 @@
-a	1
-b	2
-c	3.5
-a	1
-b	2
-c	3.5
-d	1
-e	2
-__SHARD_COUNT__	2
diff --git a/training/dtrain/hstreaming/lplp.rb b/training/dtrain/lplp.rb
index f0cd58c5..86e835e8 100755
--- a/training/dtrain/hstreaming/lplp.rb
+++ b/training/dtrain/lplp.rb
@@ -84,34 +84,28 @@ def _test()
 end
 #_test()
 
-# actually do something
+
 def usage()
-  puts "lplp.rb <l0,l1,l2,linfty,mean,median> <cut|select_k> <k|threshold> [n] < <input>"
+  puts "lplp.rb <l0,l1,l2,linfty,mean,median> <cut|select_k> <k|threshold> <#shards> < <input>"
   puts "   l0...: norms for selection"
   puts "select_k: only output top k (according to the norm of their column vector) features"
   puts "     cut: output features with weight >= threshold"
   puts "       n: if we do not have a shard count use this number for averaging"
-  exit
+  exit 1
 end
 
-if ARGV.size < 3 then usage end
+if ARGV.size < 4 then usage end
 norm_fun = method(ARGV[0].to_sym)
 type = ARGV[1]
 x = ARGV[2].to_f
-
-shard_count_key = "__SHARD_COUNT__"
+shard_count = ARGV[3].to_f
 
 STDIN.set_encoding 'utf-8'
 STDOUT.set_encoding 'utf-8'
 
 w = {}
-shard_count = 0
 while line = STDIN.gets
   key, val = line.split /\s+/
-  if key == shard_count_key
-    shard_count += 1
-    next
-  end
   if w.has_key? key
     w[key].push val.to_f
   else
@@ -119,8 +113,6 @@ while line = STDIN.gets
   end
 end
 
-if ARGV.size == 4 then shard_count = ARGV[3].to_f end
-
 if type == 'cut'
   cut(w, norm_fun, shard_count, x)
 elsif type == 'select_k'
diff --git a/training/dtrain/pairsampling.h b/training/dtrain/pairsampling.h
index 84be1efb..3f67e209 100644
--- a/training/dtrain/pairsampling.h
+++ b/training/dtrain/pairsampling.h
@@ -19,7 +19,7 @@ cmp_hyp_by_score_d(ScoredHyp a, ScoredHyp b)
 }
 
 inline void
-all_pairs(vector<ScoredHyp>* s, vector<pair<ScoredHyp,ScoredHyp> >& training, score_t threshold, unsigned max, float _unused=1)
+all_pairs(vector<ScoredHyp>* s, vector<pair<ScoredHyp,ScoredHyp> >& training, score_t threshold, unsigned max, bool misranked_only, float _unused=1)
 {
   sort(s->begin(), s->end(), cmp_hyp_by_score_d);
   unsigned sz = s->size();
@@ -27,6 +27,7 @@ all_pairs(vector<ScoredHyp>* s, vector<pair<ScoredHyp,ScoredHyp> >& training, sc
   unsigned count = 0;
   for (unsigned i = 0; i < sz-1; i++) {
     for (unsigned j = i+1; j < sz; j++) {
+      if (misranked_only && !((*s)[i].model <= (*s)[j].model)) continue;
       if (threshold > 0) {
         if (accept_pair((*s)[i].score, (*s)[j].score, threshold))
           training.push_back(make_pair((*s)[i], (*s)[j]));
@@ -51,7 +52,7 @@ all_pairs(vector<ScoredHyp>* s, vector<pair<ScoredHyp,ScoredHyp> >& training, sc
  */
 
 inline void
-partXYX(vector<ScoredHyp>* s, vector<pair<ScoredHyp,ScoredHyp> >& training, score_t threshold, unsigned max, float hi_lo)
+partXYX(vector<ScoredHyp>* s, vector<pair<ScoredHyp,ScoredHyp> >& training, score_t threshold, unsigned max, bool misranked_only, float hi_lo)
 {
   unsigned sz = s->size();
   if (sz < 2) return;
@@ -64,9 +65,7 @@ partXYX(vector<ScoredHyp>* s, vector<pair<ScoredHyp,ScoredHyp> >& training, scor
   unsigned count = 0;
   for (unsigned i = 0; i < sep_hi; i++) {
     for (unsigned j = sep_hi; j < sz; j++) {
-#ifdef DTRAIN_FASTER_PERCEPTRON
-      if ((*s)[i].model <= (*s)[j].model) {
-#endif
+      if (misranked_only && !((*s)[i].model <= (*s)[j].model)) continue;
       if (threshold > 0) {
         if (accept_pair((*s)[i].score, (*s)[j].score, threshold))
           training.push_back(make_pair((*s)[i], (*s)[j]));
@@ -78,9 +77,6 @@ partXYX(vector<ScoredHyp>* s, vector<pair<ScoredHyp,ScoredHyp> >& training, scor
         b = true;
         break;
       }
-#ifdef DTRAIN_FASTER_PERCEPTRON
-      }
-#endif
     }
     if (b) break;
   }
@@ -88,9 +84,7 @@ partXYX(vector<ScoredHyp>* s, vector<pair<ScoredHyp,ScoredHyp> >& training, scor
   while (sep_lo > 0 && (*s)[sep_lo-1].score == (*s)[sep_lo].score) --sep_lo;
   for (unsigned i = sep_hi; i < sz-sep_lo; i++) {
     for (unsigned j = sz-sep_lo; j < sz; j++) {
-#ifdef DTRAIN_FASTER_PERCEPTRON
-      if ((*s)[i].model <= (*s)[j].model) {
-#endif
+      if (misranked_only && !((*s)[i].model <= (*s)[j].model)) continue;
       if (threshold > 0) {
         if (accept_pair((*s)[i].score, (*s)[j].score, threshold))
           training.push_back(make_pair((*s)[i], (*s)[j]));
@@ -99,9 +93,6 @@ partXYX(vector<ScoredHyp>* s, vector<pair<ScoredHyp,ScoredHyp> >& training, scor
           training.push_back(make_pair((*s)[i], (*s)[j]));
       }
       if (++count == max) return;
-#ifdef DTRAIN_FASTER_PERCEPTRON
-      }
-#endif
     }
   }
 }
@@ -119,7 +110,7 @@ _PRO_cmp_pair_by_diff_d(pair<ScoredHyp,ScoredHyp> a, pair<ScoredHyp,ScoredHyp> b
   return (fabs(a.first.score - a.second.score)) > (fabs(b.first.score - b.second.score));
 }
 inline void
-PROsampling(vector<ScoredHyp>* s, vector<pair<ScoredHyp,ScoredHyp> >& training, score_t threshold, unsigned max, float _unused=1)
+PROsampling(vector<ScoredHyp>* s, vector<pair<ScoredHyp,ScoredHyp> >& training, score_t threshold, unsigned max, bool _unused=false, float _also_unused=0)
 {
   unsigned max_count = 5000, count = 0, sz = s->size();
   bool b = false;
diff --git a/training/dtrain/parallelize.rb b/training/dtrain/parallelize.rb
index eb4148f5..e661416e 100755
--- a/training/dtrain/parallelize.rb
+++ b/training/dtrain/parallelize.rb
@@ -1,80 +1,149 @@
 #!/usr/bin/env ruby
 
+require 'trollop'
 
-if ARGV.size != 5
+def usage
   STDERR.write "Usage: "
-  STDERR.write "ruby parallelize.rb <#shards> <input> <refs> <epochs> <dtrain.ini>\n"
-  exit
+  STDERR.write "ruby parallelize.rb -c <dtrain.ini> [-e <epochs=10>] [--randomize/-z] [--reshard/-y] -s <#shards|0> [-p <at once=9999>] -i <input> -r <refs> [--qsub/-q] [--dtrain_binary <path to dtrain binary>] [-l \"l2 select_k 100000\"]\n"
+  exit 1
 end
 
-cdec_dir   = '/path/to/cdec_dir'
-dtrain_bin = "#{cdec_dir}/training/dtrain/dtrain_local"
-ruby       = '/usr/bin/ruby'
-lplp_rb    = "#{cdec_dir}/training/dtrain/hstreaming/lplp.rb"
-lplp_args  = 'l2 select_k 100000'
-gzip       = '/bin/gzip'
+opts = Trollop::options do
+  opt :config, "dtrain config file", :type => :string
+  opt :epochs, "number of epochs", :type => :int, :default => 10
+  opt :lplp_args, "arguments for lplp.rb", :type => :string, :default => "l2 select_k 100000"
+  opt :randomize, "randomize shards before each epoch", :type => :bool, :short => '-z', :default => false
+  opt :reshard, "reshard after each epoch", :type => :bool, :short => '-y', :default => false
+  opt :shards, "number of shards", :type => :int
+  opt :processes_at_once, "have this number (max) running at the same time", :type => :int, :default => 9999
+  opt :input, "input", :type => :string
+  opt :references, "references", :type => :string
+  opt :qsub, "use qsub", :type => :bool, :default => false
+  opt :dtrain_binary, "path to dtrain binary", :type => :string
+end
+usage if not opts[:config]&&opts[:shards]&&opts[:input]&&opts[:references]
+
 
-num_shards = ARGV[0].to_i
-input      = ARGV[1]
-refs       = ARGV[2]
-epochs     = ARGV[3].to_i
-ini        = ARGV[4]
+dtrain_dir = File.expand_path File.dirname(__FILE__)
+if not opts[:dtrain_binary]
+  dtrain_bin = "#{dtrain_dir}/dtrain"
+else
+  dtrain_bin = opts[:dtrain_binary]
+end
+ruby       = '/usr/bin/ruby'
+lplp_rb    = "#{dtrain_dir}/lplp.rb"
+lplp_args  = opts[:lplp_args]
+cat        = '/bin/cat'
 
+ini        = opts[:config]
+epochs     = opts[:epochs]
+rand       = opts[:randomize]
+reshard    = opts[:reshard]
+predefined_shards = false
+if opts[:shards] == 0
+  predefined_shards = true
+  num_shards = 0
+else
+  num_shards = opts[:shards]
+end
+input = opts[:input]
+refs  = opts[:references]
+use_qsub       = opts[:qsub]
+shards_at_once = opts[:processes_at_once]
 
 `mkdir work`
 
-def make_shards(input, refs, num_shards)
+def make_shards(input, refs, num_shards, epoch, rand)
   lc = `wc -l #{input}`.split.first.to_i
+  index = (0..lc-1).to_a
+  index.reverse!
+  index.shuffle! if rand
   shard_sz = lc / num_shards
   leftover = lc % num_shards
   in_f = File.new input, 'r'
+  in_lines = in_f.readlines
   refs_f = File.new refs, 'r'
+  refs_lines = refs_f.readlines
   shard_in_files = []
   shard_refs_files = []
+  in_fns = []
+  refs_fns = []
   0.upto(num_shards-1) { |shard|
-    shard_in = File.new "work/shard.#{shard}.in", 'w+'
-    shard_refs = File.new "work/shard.#{shard}.refs", 'w+'
+    in_fn = "work/shard.#{shard}.#{epoch}.in"
+    shard_in = File.new in_fn, 'w+'
+    in_fns << in_fn
+    refs_fn = "work/shard.#{shard}.#{epoch}.refs"
+    shard_refs = File.new refs_fn, 'w+'
+    refs_fns << refs_fn
     0.upto(shard_sz-1) { |i|
-      shard_in.write in_f.gets
-      shard_refs.write refs_f.gets
+      j = index.pop
+      shard_in.write in_lines[j]
+      shard_refs.write refs_lines[j]
     }
     shard_in_files << shard_in
     shard_refs_files << shard_refs
   }
   while leftover > 0
-    shard_in_files[-1].write in_f.gets
-    shard_refs_files[-1].write refs_f.gets
+    j = index.pop
+    shard_in_files[-1].write in_lines[j]
+    shard_refs_files[-1].write refs_lines[j]
     leftover -= 1
   end
   (shard_in_files + shard_refs_files).each do |f| f.close end
   in_f.close
   refs_f.close
+  return [in_fns, refs_fns]
 end
 
-make_shards input, refs, num_shards
+input_files = []
+refs_files = []
+if predefined_shards
+  input_files = File.new(input).readlines.map {|i| i.strip }
+  refs_files = File.new(refs).readlines.map {|i| i.strip }
+  num_shards = input_files.size
+else
+  input_files, refs_files = make_shards input, refs, num_shards, 0, rand
+end
 
 0.upto(epochs-1) { |epoch|
+  puts "epoch #{epoch+1}"
   pids = []
   input_weights = ''
   if epoch > 0 then input_weights = "--input_weights work/weights.#{epoch-1}" end
   weights_files = []
-  0.upto(num_shards-1) { |shard|
-    pids << Kernel.fork {
-      `#{dtrain_bin} -c #{ini}\
-        --input work/shard.#{shard}.in\
-        --refs work/shard.#{shard}.refs #{input_weights}\
-        --output work/weights.#{shard}.#{epoch}\
-        &> work/out.#{shard}.#{epoch}`
+  shard = 0
+  remaining_shards = num_shards
+  while remaining_shards > 0
+    shards_at_once.times {
+      break if remaining_shards==0
+      qsub_str_start = qsub_str_end = ''
+      local_end = ''
+      if use_qsub
+        qsub_str_start = "qsub -cwd -sync y -b y -j y -o work/out.#{shard}.#{epoch} -N dtrain.#{shard}.#{epoch} \""
+        qsub_str_end = "\""
+        local_end = ''
+      else
+        local_end = "&>work/out.#{shard}.#{epoch}"
+      end
+      pids << Kernel.fork {
+        `#{qsub_str_start}#{dtrain_bin} -c #{ini}\
+          --input #{input_files[shard]}\
+          --refs #{refs_files[shard]} #{input_weights}\
+          --output work/weights.#{shard}.#{epoch}#{qsub_str_end} #{local_end}`
+      }
+      weights_files << "work/weights.#{shard}.#{epoch}"
+      shard += 1
+      remaining_shards -= 1
     }
-    weights_files << "work/weights.#{shard}.#{epoch}"
-  }
-  pids.each { |pid| Process.wait(pid) }
-  cat = File.new('work/weights_cat', 'w+')
-  weights_files.each { |f| cat.write File.new(f, 'r').read }
-  cat.close
-  `#{ruby} #{lplp_rb} #{lplp_args} #{num_shards} < work/weights_cat &> work/weights.#{epoch}`
+    pids.each { |pid| Process.wait(pid) }
+    pids.clear
+  end
+  `#{cat} work/weights.*.#{epoch} > work/weights_cat`
+  `#{ruby} #{lplp_rb} #{lplp_args} #{num_shards} < work/weights_cat > work/weights.#{epoch}`
+  if rand and reshard and epoch+1!=epochs
+    input_files, refs_files = make_shards input, refs, num_shards, epoch+1, rand
+  end
 }
 
 `rm work/weights_cat`
-`#{gzip} work/*`
 
diff --git a/training/dtrain/score.cc b/training/dtrain/score.cc
index 34fc86a9..96d6e10a 100644
--- a/training/dtrain/score.cc
+++ b/training/dtrain/score.cc
@@ -49,7 +49,7 @@ BleuScorer::Score(vector<WordID>& hyp, vector<WordID>& ref,
  *        for Machine Translation"
  * (Lin & Och '04)
  *
- * NOTE: 0 iff no 1gram match
+ * NOTE: 0 iff no 1gram match ('grounded')
  */
 score_t
 StupidBleuScorer::Score(vector<WordID>& hyp, vector<WordID>& ref,
@@ -74,6 +74,35 @@ StupidBleuScorer::Score(vector<WordID>& hyp, vector<WordID>& ref,
 }
 
 /*
+ * fixed 'stupid' bleu
+ *
+ * as in "Optimizing for Sentence-Level BLEU+1
+ *        Yields Short Translations"
+ * (Nakov et al. '12)
+ */
+score_t
+FixedStupidBleuScorer::Score(vector<WordID>& hyp, vector<WordID>& ref,
+                        const unsigned /*rank*/, const unsigned /*src_len*/)
+{
+  unsigned hyp_len = hyp.size(), ref_len = ref.size();
+  if (hyp_len == 0 || ref_len == 0) return 0.;
+  NgramCounts counts = make_ngram_counts(hyp, ref, N_);
+  unsigned M = N_;
+  vector<score_t> v = w_;
+  if (ref_len < N_) {
+    M = ref_len;
+    for (unsigned i = 0; i < M; i++) v[i] = 1/((score_t)M);
+  }
+  score_t sum = 0, add = 0;
+  for (unsigned i = 0; i < M; i++) {
+    if (i == 0 && (counts.sum_[i] == 0 || counts.clipped_[i] == 0)) return 0.;
+    if (i == 1) add = 1;
+    sum += v[i] * log(((score_t)counts.clipped_[i] + add)/((counts.sum_[i] + add)));
+  }
+  return  brevity_penalty(hyp_len, ref_len+1) * exp(sum); // <- fix
+}
+
+/*
  * smooth bleu
  *
  * as in "An End-to-End Discriminative Approach
diff --git a/training/dtrain/score.h b/training/dtrain/score.h
index f317c903..bddaa071 100644
--- a/training/dtrain/score.h
+++ b/training/dtrain/score.h
@@ -148,6 +148,11 @@ struct StupidBleuScorer : public LocalScorer
   score_t Score(vector<WordID>& hyp, vector<WordID>& ref, const unsigned /*rank*/, const unsigned /*src_len*/);
 };
 
+struct FixedStupidBleuScorer : public LocalScorer
+{
+  score_t Score(vector<WordID>& hyp, vector<WordID>& ref, const unsigned /*rank*/, const unsigned /*src_len*/);
+};
+
 struct SmoothBleuScorer : public LocalScorer
 {
   score_t Score(vector<WordID>& hyp, vector<WordID>& ref, const unsigned /*rank*/, const unsigned /*src_len*/);
diff --git a/training/dtrain/test/example/README b/training/dtrain/test/example/README
deleted file mode 100644
index 6937b11b..00000000
--- a/training/dtrain/test/example/README
+++ /dev/null
@@ -1,8 +0,0 @@
-Small example of input format for distributed training.
-Call dtrain from cdec/dtrain/ with ./dtrain -c test/example/dtrain.ini .
-
-For this to work, undef 'DTRAIN_LOCAL' in dtrain.h
-and recompile.
-
-Data is here: http://simianer.de/#dtrain
-
diff --git a/training/dtrain/test/example/expected-output b/training/dtrain/test/example/expected-output
deleted file mode 100644
index 05326763..00000000
--- a/training/dtrain/test/example/expected-output
+++ /dev/null
@@ -1,89 +0,0 @@
-                cdec cfg 'test/example/cdec.ini'
-Loading the LM will be faster if you build a binary file.
-Reading test/example/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
-****************************************************************************************************
-  Example feature: Shape_S00000_T00000
-Seeding random number sequence to 2912000813
-
-dtrain
-Parameters:
-                       k 100
-                       N 4
-                       T 2
-                 scorer 'stupid_bleu'
-             sample from 'kbest'
-                  filter 'uniq'
-           learning rate 1
-                   gamma 0
-             loss margin 0
-                   pairs 'XYX'
-                   hi lo 0.1
-          pair threshold 0
-          select weights 'VOID'
-                  l1 reg 0 'none'
-               max pairs 4294967295
-                cdec cfg 'test/example/cdec.ini'
-                   input 'test/example/nc-wmt11.1k.gz'
-                  output '-'
-              stop_after 10
-(a dot represents 10 inputs)
-Iteration #1 of 2.
- . 10
-Stopping after 10 input sentences.
-WEIGHTS
-              Glue = -637
-       WordPenalty = +1064
-     LanguageModel = +1175.3
- LanguageModel_OOV = -1437
-     PhraseModel_0 = +1935.6
-     PhraseModel_1 = +2499.3
-     PhraseModel_2 = +964.96
-     PhraseModel_3 = +1410.8
-     PhraseModel_4 = -5977.9
-     PhraseModel_5 = +522
-     PhraseModel_6 = +1089
-       PassThrough = -1308
-        ---
-       1best avg score: 0.16963 (+0.16963)
- 1best avg model score: 64485 (+64485)
-           avg # pairs: 1494.4
-        avg # rank err: 702.6
-     avg # margin viol: 0
-    non0 feature count: 528
-           avg list sz: 85.7
-           avg f count: 102.75
-(time 0.083 min, 0.5 s/S)
-
-Iteration #2 of 2.
- . 10
-WEIGHTS
-              Glue = -1196
-       WordPenalty = +809.52
-     LanguageModel = +3112.1
- LanguageModel_OOV = -1464
-     PhraseModel_0 = +3895.5
-     PhraseModel_1 = +4683.4
-     PhraseModel_2 = +1092.8
-     PhraseModel_3 = +1079.6
-     PhraseModel_4 = -6827.7
-     PhraseModel_5 = -888
-     PhraseModel_6 = +142
-       PassThrough = -1335
-        ---
-       1best avg score: 0.277 (+0.10736)
- 1best avg model score: -3110.5 (-67595)
-           avg # pairs: 1144.2
-        avg # rank err: 529.1
-     avg # margin viol: 0
-    non0 feature count: 859
-           avg list sz: 74.9
-           avg f count: 112.84
-(time 0.067 min, 0.4 s/S)
-
-Writing weights file to '-' ...
-done
-
----
-Best iteration: 2 [SCORE 'stupid_bleu'=0.277].
-This took 0.15 min.
diff --git a/training/dtrain/test/parallelize/in b/training/dtrain/test/parallelize/in
deleted file mode 100644
index a312809f..00000000
--- a/training/dtrain/test/parallelize/in
+++ /dev/null
@@ -1,10 +0,0 @@
-<seg grammar="/stor/dat/wmt12/dev/newstest2010/g/grammar.out.0.gz" id="0">barack obama erhält als vierter us @-@ präsident den frieden nobelpreis</seg>
-<seg grammar="/stor/dat/wmt12/dev/newstest2010/g/grammar.out.1.gz" id="1">der amerikanische präsident barack obama kommt für 26 stunden nach oslo , norwegen , um hier als vierter us @-@ präsident in der geschichte den frieden nobelpreis entgegen zunehmen .</seg>
-<seg grammar="/stor/dat/wmt12/dev/newstest2010/g/grammar.out.2.gz" id="2">darüber hinaus erhält er das diplom sowie die medaille und einen scheck über 1,4 mio. dollar für seine außer gewöhnlichen bestrebungen um die intensivierung der welt diplomatie und zusammen arbeit unter den völkern .</seg>
-<seg grammar="/stor/dat/wmt12/dev/newstest2010/g/grammar.out.3.gz" id="3">der chef des weißen hauses kommt morgen zusammen mit seiner frau michelle in der nordwegischen metropole an und wird die ganze zeit beschäftigt sein .</seg>
-<seg grammar="/stor/dat/wmt12/dev/newstest2010/g/grammar.out.4.gz" id="4">zunächst stattet er dem nobel @-@ institut einen besuch ab , wo er überhaupt zum ersten mal mit den fünf ausschuss mitglieder zusammen trifft , die ihn im oktober aus 172 leuten und 33 organisationen gewählt haben .</seg>
-<seg grammar="/stor/dat/wmt12/dev/newstest2010/g/grammar.out.5.gz" id="5">das präsidenten paar hat danach ein treffen mit dem norwegischen könig harald v. und königin sonja eingeplant .</seg>
-<seg grammar="/stor/dat/wmt12/dev/newstest2010/g/grammar.out.6.gz" id="6">nachmittags erreicht dann der besuch seinen höhepunkt mit der zeremonie , bei der obama den prestige preis übernimmt .</seg>
-<seg grammar="/stor/dat/wmt12/dev/newstest2010/g/grammar.out.7.gz" id="7">diesen erhält er als der vierte us @-@ präsident , aber erst als der dritte , der den preis direkt im amt entgegen nimmt .</seg>
-<seg grammar="/stor/dat/wmt12/dev/newstest2010/g/grammar.out.8.gz" id="8">das weiße haus avisierte schon , dass obama bei der übernahme des preises über den afghanistan krieg sprechen wird .</seg>
-<seg grammar="/stor/dat/wmt12/dev/newstest2010/g/grammar.out.9.gz" id="9">der präsident will diesem thema nicht ausweichen , weil er weiß , dass er den preis als ein präsident übernimmt , der zur zeit krieg in zwei ländern führt .</seg>
diff --git a/training/dtrain/test/parallelize/refs b/training/dtrain/test/parallelize/refs
deleted file mode 100644
index 4d3128cb..00000000
--- a/training/dtrain/test/parallelize/refs
+++ /dev/null
@@ -1,10 +0,0 @@
-barack obama becomes the fourth american president to receive the nobel peace prize
-the american president barack obama will fly into oslo , norway for 26 hours to receive the nobel peace prize , the fourth american president in history to do so .
-he will receive a diploma , medal and cheque for 1.4 million dollars for his exceptional efforts to improve global diplomacy and encourage international cooperation , amongst other things .
-the head of the white house will be flying into the norwegian city in the morning with his wife michelle and will have a busy schedule .
-first , he will visit the nobel institute , where he will have his first meeting with the five committee members who selected him from 172 people and 33 organisations .
-the presidential couple then has a meeting scheduled with king harald v and queen sonja of norway .
-then , in the afternoon , the visit will culminate in a grand ceremony , at which obama will receive the prestigious award .
-he will be the fourth american president to be awarded the prize , and only the third to have received it while actually in office .
-the white house has stated that , when he accepts the prize , obama will speak about the war in afghanistan .
-the president does not want to skirt around this topic , as he realises that he is accepting the prize as a president whose country is currently at war in two countries .
diff --git a/training/dtrain/test/parallelize/test/cdec.ini b/training/dtrain/test/parallelize/test/cdec.ini
deleted file mode 100644
index 72e99dc5..00000000
--- a/training/dtrain/test/parallelize/test/cdec.ini
+++ /dev/null
@@ -1,22 +0,0 @@
-formalism=scfg
-add_pass_through_rules=true
-intersection_strategy=cube_pruning
-cubepruning_pop_limit=200
-scfg_max_span_limit=15
-feature_function=WordPenalty
-feature_function=KLanguageModel /stor/dat/wmt12/en/news_only/m/wmt12.news.en.3.kenv5
-#feature_function=ArityPenalty
-#feature_function=CMR2008ReorderingFeatures
-#feature_function=Dwarf
-#feature_function=InputIndicator
-#feature_function=LexNullJump
-#feature_function=NewJump
-#feature_function=NgramFeatures
-#feature_function=NonLatinCount
-#feature_function=OutputIndicator
-#feature_function=RuleIdentityFeatures
-#feature_function=RuleNgramFeatures
-#feature_function=RuleShape
-#feature_function=SourceSpanSizeFeatures
-#feature_function=SourceWordPenalty
-#feature_function=SpanFeatures
diff --git a/training/dtrain/test/parallelize/test/dtrain.ini b/training/dtrain/test/parallelize/test/dtrain.ini
deleted file mode 100644
index 03f9d240..00000000
--- a/training/dtrain/test/parallelize/test/dtrain.ini
+++ /dev/null
@@ -1,15 +0,0 @@
-k=100
-N=4
-learning_rate=0.0001
-gamma=0
-loss_margin=0
-epochs=1
-scorer=stupid_bleu
-sample_from=kbest
-filter=uniq
-pair_sampling=XYX
-hi_lo=0.1
-select_weights=last
-print_weights=Glue WordPenalty LanguageModel LanguageModel_OOV PhraseModel_0 PhraseModel_1 PhraseModel_2 PhraseModel_3 PhraseModel_4 PhraseModel_5 PhraseModel_6 PassThrough
-tmp=/tmp
-decoder_config=cdec.ini
diff --git a/training/dtrain/test/parallelize/test/in b/training/dtrain/test/parallelize/test/in
deleted file mode 100644
index a312809f..00000000
--- a/training/dtrain/test/parallelize/test/in
+++ /dev/null
@@ -1,10 +0,0 @@
-<seg grammar="/stor/dat/wmt12/dev/newstest2010/g/grammar.out.0.gz" id="0">barack obama erhält als vierter us @-@ präsident den frieden nobelpreis</seg>
-<seg grammar="/stor/dat/wmt12/dev/newstest2010/g/grammar.out.1.gz" id="1">der amerikanische präsident barack obama kommt für 26 stunden nach oslo , norwegen , um hier als vierter us @-@ präsident in der geschichte den frieden nobelpreis entgegen zunehmen .</seg>
-<seg grammar="/stor/dat/wmt12/dev/newstest2010/g/grammar.out.2.gz" id="2">darüber hinaus erhält er das diplom sowie die medaille und einen scheck über 1,4 mio. dollar für seine außer gewöhnlichen bestrebungen um die intensivierung der welt diplomatie und zusammen arbeit unter den völkern .</seg>
-<seg grammar="/stor/dat/wmt12/dev/newstest2010/g/grammar.out.3.gz" id="3">der chef des weißen hauses kommt morgen zusammen mit seiner frau michelle in der nordwegischen metropole an und wird die ganze zeit beschäftigt sein .</seg>
-<seg grammar="/stor/dat/wmt12/dev/newstest2010/g/grammar.out.4.gz" id="4">zunächst stattet er dem nobel @-@ institut einen besuch ab , wo er überhaupt zum ersten mal mit den fünf ausschuss mitglieder zusammen trifft , die ihn im oktober aus 172 leuten und 33 organisationen gewählt haben .</seg>
-<seg grammar="/stor/dat/wmt12/dev/newstest2010/g/grammar.out.5.gz" id="5">das präsidenten paar hat danach ein treffen mit dem norwegischen könig harald v. und königin sonja eingeplant .</seg>
-<seg grammar="/stor/dat/wmt12/dev/newstest2010/g/grammar.out.6.gz" id="6">nachmittags erreicht dann der besuch seinen höhepunkt mit der zeremonie , bei der obama den prestige preis übernimmt .</seg>
-<seg grammar="/stor/dat/wmt12/dev/newstest2010/g/grammar.out.7.gz" id="7">diesen erhält er als der vierte us @-@ präsident , aber erst als der dritte , der den preis direkt im amt entgegen nimmt .</seg>
-<seg grammar="/stor/dat/wmt12/dev/newstest2010/g/grammar.out.8.gz" id="8">das weiße haus avisierte schon , dass obama bei der übernahme des preises über den afghanistan krieg sprechen wird .</seg>
-<seg grammar="/stor/dat/wmt12/dev/newstest2010/g/grammar.out.9.gz" id="9">der präsident will diesem thema nicht ausweichen , weil er weiß , dass er den preis als ein präsident übernimmt , der zur zeit krieg in zwei ländern führt .</seg>
diff --git a/training/dtrain/test/parallelize/test/refs b/training/dtrain/test/parallelize/test/refs
deleted file mode 100644
index 4d3128cb..00000000
--- a/training/dtrain/test/parallelize/test/refs
+++ /dev/null
@@ -1,10 +0,0 @@
-barack obama becomes the fourth american president to receive the nobel peace prize
-the american president barack obama will fly into oslo , norway for 26 hours to receive the nobel peace prize , the fourth american president in history to do so .
-he will receive a diploma , medal and cheque for 1.4 million dollars for his exceptional efforts to improve global diplomacy and encourage international cooperation , amongst other things .
-the head of the white house will be flying into the norwegian city in the morning with his wife michelle and will have a busy schedule .
-first , he will visit the nobel institute , where he will have his first meeting with the five committee members who selected him from 172 people and 33 organisations .
-the presidential couple then has a meeting scheduled with king harald v and queen sonja of norway .
-then , in the afternoon , the visit will culminate in a grand ceremony , at which obama will receive the prestigious award .
-he will be the fourth american president to be awarded the prize , and only the third to have received it while actually in office .
-the white house has stated that , when he accepts the prize , obama will speak about the war in afghanistan .
-the president does not want to skirt around this topic , as he realises that he is accepting the prize as a president whose country is currently at war in two countries .
diff --git a/training/dtrain/test/toy/input b/training/dtrain/test/toy/input
deleted file mode 100644
index 4d10a9ea..00000000
--- a/training/dtrain/test/toy/input
+++ /dev/null
@@ -1,2 +0,0 @@
-0	ich sah ein kleines haus	i saw a little house	[S] ||| [NP,1] [VP,2] ||| [1] [2] ||| logp=0	[NP] ||| ich ||| i ||| logp=0	[NP] ||| ein [NN,1] ||| a [1] ||| logp=0	[NN] ||| [JJ,1] haus ||| [1] house ||| logp=0 house_rule=1	[NN] ||| [JJ,1] haus ||| [1] shell ||| logp=0 shell_rule=1	[JJ] ||| kleines ||| small ||| logp=0 small_rule=1	[JJ] ||| kleines ||| little ||| logp=0 little_rule=1	[JJ] ||| grosses ||| big ||| logp=0	[JJ] ||| grosses ||| large ||| logp=0	[VP] ||| [V,1] [NP,2] ||| [1] [2] ||| logp=0	[V] ||| sah ||| saw ||| logp=0	[V] ||| fand ||| found ||| logp=0
-1	ich fand ein kleines haus	i found a little house	[S] ||| [NP,1] [VP,2] ||| [1] [2] ||| logp=0	[NP] ||| ich ||| i ||| logp=0	[NP] ||| ein [NN,1] ||| a [1] ||| logp=0	[NN] ||| [JJ,1] haus ||| [1] house ||| logp=0 house_rule=1	[NN] ||| [JJ,1] haus ||| [1] shell ||| logp=0 shell_rule=1	[JJ] ||| kleines ||| small ||| logp=0 small_rule=1	[JJ] ||| kleines ||| little ||| logp=0 little_rule=1	[JJ] ||| grosses ||| big ||| logp=0	[JJ] ||| grosses ||| large ||| logp=0	[VP] ||| [V,1] [NP,2] ||| [1] [2] ||| logp=0	[V] ||| sah ||| saw ||| logp=0	[V] ||| fand ||| found ||| logp=0