diff --git a/python/paddle/v2/dataset/common.py b/python/paddle/v2/dataset/common.py
index a5ffe25a116e9be039bdebaaaad435685e23d372..fcf4437ffaf329f52cc5bc997eff45dee200873c 100644
--- a/python/paddle/v2/dataset/common.py
+++ b/python/paddle/v2/dataset/common.py
@@ -32,3 +32,10 @@ def download(url, module_name, md5sum):
             shutil.copyfileobj(r.raw, f)
 
     return filename
+
+
+def dict_add(a_dict, ele):
+    if ele in a_dict:
+        a_dict[ele] += 1
+    else:
+        a_dict[ele] = 1
diff --git a/python/paddle/v2/dataset/imdb.py b/python/paddle/v2/dataset/imdb.py
index 7a191cb583d4ced8e92354f4158e646d57b601bd..433e37380f840f5b7ff619a5f64b99d2ad724b17 100644
--- a/python/paddle/v2/dataset/imdb.py
+++ b/python/paddle/v2/dataset/imdb.py
@@ -17,3 +17,104 @@
 """
 IMDB dataset: http://ai.stanford.edu/%7Eamaas/data/sentiment/aclImdb_v1.tar.gz
 """
+import paddle.v2.dataset.common
+import tarfile
+import Queue
+import re
+import string
+import threading
+
+__all__ = ['build_dict', 'train', 'test']
+
+URL = 'http://ai.stanford.edu/%7Eamaas/data/sentiment/aclImdb_v1.tar.gz'
+MD5 = '7c2ac02c03563afcf9b574c7e56c153a'
+
+
+# Read files that match pattern.  Tokenize and yield each file.
+def tokenize(pattern):
+    with tarfile.open(paddle.v2.dataset.common.download(URL, 'imdb',
+                                                        MD5)) as tarf:
+        # Note that we should use tarfile.next(), which does
+        # sequential access of member files, other than
+        # tarfile.extractfile, which does random access and might
+        # destroy hard disks.
+        tf = tarf.next()
+        while tf != None:
+            if bool(pattern.match(tf.name)):
+                # newline and punctuations removal and ad-hoc tokenization.
+                yield tarf.extractfile(tf).read().rstrip("\n\r").translate(
+                    None, string.punctuation).lower().split()
+            tf = tarf.next()
+
+
+def build_dict(pattern, cutoff):
+    word_freq = {}
+    for doc in tokenize(pattern):
+        for word in doc:
+            paddle.v2.dataset.common.dict_add(word_freq, word)
+
+    # Not sure if we should prune less-frequent words here.
+    word_freq = filter(lambda x: x[1] > cutoff, word_freq.items())
+
+    dictionary = sorted(word_freq, key=lambda x: (-x[1], x[0]))
+    words, _ = list(zip(*dictionary))
+    word_idx = dict(zip(words, xrange(len(words))))
+    word_idx['<unk>'] = len(words)
+    return word_idx
+
+
+def reader_creator(pos_pattern, neg_pattern, word_idx, buffer_size):
+    UNK = word_idx['<unk>']
+
+    qs = [Queue.Queue(maxsize=buffer_size), Queue.Queue(maxsize=buffer_size)]
+
+    def load(pattern, queue):
+        for doc in tokenize(pattern):
+            queue.put(doc)
+        queue.put(None)
+
+    def reader():
+        # Creates two threads that loads positive and negative samples
+        # into qs.
+        t0 = threading.Thread(
+            target=load, args=(
+                pos_pattern,
+                qs[0], ))
+        t0.daemon = True
+        t0.start()
+
+        t1 = threading.Thread(
+            target=load, args=(
+                neg_pattern,
+                qs[1], ))
+        t1.daemon = True
+        t1.start()
+
+        # Read alternatively from qs[0] and qs[1].
+        i = 0
+        doc = qs[i].get()
+        while doc != None:
+            yield [word_idx.get(w, UNK) for w in doc], i % 2
+            i += 1
+            doc = qs[i % 2].get()
+
+        # If any queue is empty, reads from the other queue.
+        i += 1
+        doc = qs[i % 2].get()
+        while doc != None:
+            yield [word_idx.get(w, UNK) for w in doc], i % 2
+            doc = qs[i % 2].get()
+
+    return reader()
+
+
+def train(word_idx):
+    return reader_creator(
+        re.compile("aclImdb/train/pos/.*\.txt$"),
+        re.compile("aclImdb/train/neg/.*\.txt$"), word_idx, 1000)
+
+
+def test(word_idx):
+    return reader_creator(
+        re.compile("aclImdb/test/pos/.*\.txt$"),
+        re.compile("aclImdb/test/neg/.*\.txt$"), word_idx, 1000)
diff --git a/python/paddle/v2/dataset/tests/imdb_test.py b/python/paddle/v2/dataset/tests/imdb_test.py
index 5cfb6203e78ec14f2cc5474876a6ed7de7d491e9..e887af16634d2db04b8cf5fa0269a69991d8baac 100644
--- a/python/paddle/v2/dataset/tests/imdb_test.py
+++ b/python/paddle/v2/dataset/tests/imdb_test.py
@@ -1,12 +1,43 @@
-import paddle.v2.dataset.common
-import tarfile
+import paddle.v2.dataset.imdb
+import unittest
+import re
 
-URL = 'http://ai.stanford.edu/%7Eamaas/data/sentiment/aclImdb_v1.tar.gz'
-MD5 = '7c2ac02c03563afcf9b574c7e56c153a'
+TRAIN_POS_PATTERN = re.compile("aclImdb/train/pos/.*\.txt$")
+TRAIN_NEG_PATTERN = re.compile("aclImdb/train/neg/.*\.txt$")
+TRAIN_PATTERN = re.compile("aclImdb/train/.*\.txt$")
 
-tarf = tarfile.open(paddle.v2.dataset.common.download(URL, 'imdb', MD5))
+TEST_POS_PATTERN = re.compile("aclImdb/test/pos/.*\.txt$")
+TEST_NEG_PATTERN = re.compile("aclImdb/test/neg/.*\.txt$")
+TEST_PATTERN = re.compile("aclImdb/test/.*\.txt$")
 
-tf = tarf.next()
-while tf != None:
-    print tf.name
-    tf = tarf.next()
+
+class TestIMDB(unittest.TestCase):
+    word_idx = None
+
+    def test_build_dict(self):
+        if self.word_idx == None:
+            self.word_idx = paddle.v2.dataset.imdb.build_dict(TRAIN_PATTERN,
+                                                              150)
+
+        self.assertEqual(len(self.word_idx), 7036)
+
+    def check_dataset(self, dataset, expected_size):
+        if self.word_idx == None:
+            self.word_idx = paddle.v2.dataset.imdb.build_dict(TRAIN_PATTERN,
+                                                              150)
+
+        sum = 0
+        for l in dataset(self.word_idx):
+            self.assertEqual(l[1], sum % 2)
+            sum += 1
+        self.assertEqual(sum, expected_size)
+
+    def test_train(self):
+        self.check_dataset(paddle.v2.dataset.imdb.train, 25000)
+
+    def test_test(self):
+        self.check_dataset(paddle.v2.dataset.imdb.test, 25000)
+
+
+if __name__ == '__main__':
+    unittest.main()