diff --git a/README.md b/README.md
index 5ca734b..c3705b7 100644
--- a/README.md
+++ b/README.md
@@ -10,15 +10,15 @@ Execution
 $ . configure.sh
 $ make && sudo make install
 $ lua test/test.lua
-<s>	<s>	=>	the
-<s>	the	=>	u.s.
-the	u.s.	=>	centers
-u.s.	centers	=>	for
-centers	for	=>	disease
-for	disease	=>	control
-disease	control	=>	and
-control	and	=>	prevention
-and	prevention	=>	in
-prevention	in	=>	atlanta
+<s> <s> => the
+<s> the => u.s.
+the u.s. => centers
+u.s. centers => for
+centers for => disease
+for disease => control
+disease control => and
+control and => prevention
+and prevention => in
+prevention in => atlanta
 ...
 ```
diff --git a/packages/nnlm.dataset/c_src/nnlm_dataset.cc b/packages/nnlm.dataset/c_src/nnlm_dataset.cc
index 3a3d261..ad2fcb1 100644
--- a/packages/nnlm.dataset/c_src/nnlm_dataset.cc
+++ b/packages/nnlm.dataset/c_src/nnlm_dataset.cc
@@ -79,7 +79,7 @@ namespace LanguageModels {
       if (length == 1) return tk; // FIXME: refactor this code
       (*pat)[i] = tk;
     }
-    return pat.release();
+    return pat.weakRelease();
   }
 
   Token *NNLMDataSetToken::getPatternBunch(const int *indexes,
@@ -111,7 +111,7 @@ namespace LanguageModels {
       if (length == 1) return tk; // FIXME: refactor this code
       (*pat)[i] = tk;
     }
-    return pat.release();
+    return pat.weakRelease();
   }
 
   void NNLMDataSetToken::putPattern(int index, Token *pat) {
diff --git a/packages/nnlm.dataset/lua_src/nnlm_dataset.lua b/packages/nnlm.dataset/lua_src/nnlm_dataset.lua
new file mode 100644
index 0000000..4eb6e03
--- /dev/null
+++ b/packages/nnlm.dataset/lua_src/nnlm_dataset.lua
@@ -0,0 +1,25 @@
+nnlm = nnlm or {}
+nnlm.dataset = nnlm.dataset or {}
+nnlm.dataset.both = function(t)
+  local params = get_table_fields({
+      corpora = { isa_match=nnlm.corpora, mandatory=true },
+      order = { type_match="number", mandatory=true },
+      initial_word = { type_match="string", mandatory=true, default="<s>" },
+      final_word = { type_match="string", mandatory=true, default="</s>" },
+                                  }, t)
+  local in_ds = nnlm.dataset{
+    offset       = -params.order + 1,
+    length       =  params.order - 1,
+    corpora      = params.corpora,
+    initial_word = params.initial_word,
+    final_word   = params.final_word,
+  }
+  local out_ds = nnlm.dataset{
+    offset       = 0,
+    length       = 1,
+    corpora      = params.corpora,
+    initial_word = params.initial_word,
+    final_word   = params.final_word,
+  }
+  return in_ds,out_ds
+end
diff --git a/test/test.lua b/test/test.lua
index f37f2a5..e122897 100644
--- a/test/test.lua
+++ b/test/test.lua
@@ -1,28 +1,37 @@
+require "aprilann" -- it is not necessary but its clear
 require "nnlmutils"
 
 local dir     = arg[0]:dirname()
 local lex     = lexClass.load(io.open(dir.."/voc.398"))
 local corpora = nnlm.corpora(dir.."/sample.txt",lex.cobj,"<unk>")
-local in_ds   = nnlm.dataset{
+local in_ds,out_ds = nnlm.dataset.both{
   corpora = corpora,
-  length  =  2, -- bigrams
-  offset  = -2, -- input requires negative offset
-  initial_word = "<s>",
-  final_word   = "</s>",
-}
-local out_ds   = nnlm.dataset{
-  corpora = corpora,
-  length  = 1, -- unigram
-  offset  = 0, -- output shouldn't requires offset
-  initial_word = "<s>",
-  final_word   = "</s>",
+  order   =  3, -- trigrams
 }
 
 local function words_of(tk)
-  return lex:getWordFromWordId(select(2,tk:to_dense():max()))
+  local idx = select(2,tk:to_dense():max(2)):squeeze()
+  local t = {}
+  for i=1,#idx do
+    t[i] = lex:getWordFromWordId(idx[i])
+  end
+  return t
 end
 
 for i,in_pat in in_ds:patterns() do
   local out_pat = out_ds:getPattern(i)
-  print(words_of(in_pat[1]), words_of(in_pat[2]), "=>", words_of(out_pat))
+  print(iterator.zip(iterator(words_of(in_pat[1])),
+                     iterator(words_of(in_pat[2])),
+                     iterator.duplicate("=>"),
+                     iterator(words_of(out_pat))):concat(" ", "\n"))
 end
+
+print("*******************************************************************")
+
+local in_pat  = in_ds:getPatternBunch(iterator.range(in_ds:numPatterns()):table())
+local out_pat = out_ds:getPatternBunch(iterator.range(in_ds:numPatterns()):table())
+
+print(iterator.zip(iterator(words_of(in_pat[1])),
+                   iterator(words_of(in_pat[2])),
+                   iterator.duplicate("=>"),
+                   iterator(words_of(out_pat))):concat(" ", "\n"))