diff --git a/examples/run_bert_classifier.py b/examples/run_bert_classifier.py
index 6f3d26cee1..27b8e6165d 100644
--- a/examples/run_bert_classifier.py
+++ b/examples/run_bert_classifier.py
@@ -309,14 +309,7 @@ def main():
 
                 # define a new function to compute loss values for both output_modes
                 ouputs = model(input_ids, token_type_ids=segment_ids, attention_mask=input_mask, labels=label_ids)
-                loss = 
-
-                if output_mode == "classification":
-                    loss_fct = CrossEntropyLoss()
-                    loss = loss_fct(logits.view(-1, num_labels), label_ids.view(-1))
-                elif output_mode == "regression":
-                    loss_fct = MSELoss()
-                    loss = loss_fct(logits.view(-1), label_ids.view(-1))
+                loss = ouputs[0]
 
                 if n_gpu > 1:
                     loss = loss.mean() # mean() to average on multi-gpu.
@@ -423,15 +416,8 @@ def main():
             label_ids = label_ids.to(device)
 
             with torch.no_grad():
-                logits = model(input_ids, token_type_ids=segment_ids, attention_mask=input_mask)
-
-            # create eval loss and other metric required by the task
-            if output_mode == "classification":
-                loss_fct = CrossEntropyLoss()
-                tmp_eval_loss = loss_fct(logits.view(-1, num_labels), label_ids.view(-1))
-            elif output_mode == "regression":
-                loss_fct = MSELoss()
-                tmp_eval_loss = loss_fct(logits.view(-1), label_ids.view(-1))
+                outputs = model(input_ids, token_type_ids=segment_ids, attention_mask=input_mask, labels=label_ids)
+                tmp_eval_loss, logits = outputs[:2]
 
             eval_loss += tmp_eval_loss.mean().item()
             nb_eval_steps += 1
diff --git a/examples/run_glue.py b/examples/run_glue.py
new file mode 100644
index 0000000000..da1e8d8123
--- /dev/null
+++ b/examples/run_glue.py
@@ -0,0 +1,401 @@
+# coding=utf-8
+# Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team.
+# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""BERT finetuning runner."""
+
+from __future__ import absolute_import, division, print_function
+
+import argparse
+import logging
+import os
+import sys
+import random
+from tqdm import tqdm, trange
+
+import numpy as np
+
+import torch
+from torch.utils.data import (DataLoader, RandomSampler, SequentialSampler,
+                              TensorDataset)
+from torch.utils.data.distributed import DistributedSampler
+from torch.nn import CrossEntropyLoss, MSELoss
+
+from tensorboardX import SummaryWriter
+
+from pytorch_transformers import WEIGHTS_NAME, CONFIG_NAME
+from pytorch_transformers.modeling_bert import BertForSequenceClassification
+from pytorch_transformers.tokenization_bert import BertTokenizer
+from pytorch_transformers.optimization import BertAdam, WarmupLinearSchedule
+
+from utils_glue import processors, output_modes, convert_examples_to_features, compute_metrics
+
+
+logger = logging.getLogger(__name__)
+
+
+def main():
+    parser = argparse.ArgumentParser()
+
+    ## Required parameters
+    parser.add_argument("--data_dir", default=None, type=str, required=True,
+                        help="The input data dir. Should contain the .tsv files (or other data files) for the task.")
+    parser.add_argument("--bert_model", default=None, type=str, required=True,
+                        help="Bert pre-trained model selected in the list: bert-base-uncased, "
+                        "bert-large-uncased, bert-base-cased, bert-large-cased, bert-base-multilingual-uncased, "
+                        "bert-base-multilingual-cased, bert-base-chinese.")
+    parser.add_argument("--task_name", default=None, type=str, required=True,
+                        help="The name of the task to train.")
+    parser.add_argument("--output_dir", default=None, type=str, required=True,
+                        help="The output directory where the model predictions and checkpoints will be written.")
+
+    ## Other parameters
+    parser.add_argument("--cache_dir", default="", type=str,
+                        help="Where do you want to store the pre-trained models downloaded from s3")
+    parser.add_argument("--max_seq_length", default=128, type=int,
+                        help="The maximum total input sequence length after WordPiece tokenization. \n"
+                             "Sequences longer than this will be truncated, and sequences shorter \n"
+                             "than this will be padded.")
+    parser.add_argument("--do_train", action='store_true',
+                        help="Whether to run training.")
+    parser.add_argument("--do_eval", action='store_true',
+                        help="Whether to run eval on the dev set.")
+    parser.add_argument("--do_lower_case", action='store_true',
+                        help="Set this flag if you are using an uncased model.")
+    parser.add_argument("--train_batch_size", default=32, type=int,
+                        help="Total batch size for training.")
+    parser.add_argument("--eval_batch_size", default=8, type=int,
+                        help="Total batch size for eval.")
+    parser.add_argument('--gradient_accumulation_steps', type=int, default=1,
+                        help="Number of updates steps to accumulate before performing a backward/update pass.")
+    parser.add_argument("--learning_rate", default=5e-5, type=float,
+                        help="The initial learning rate for Adam.")
+    parser.add_argument("--num_train_epochs", default=3.0, type=float,
+                        help="Total number of training epochs to perform.")
+    parser.add_argument("--warmup_proportion", default=0.1, type=float,
+                        help="Proportion of training to perform linear learning rate warmup for. "
+                             "E.g., 0.1 = 10%% of training.")
+    parser.add_argument("--no_cuda", action='store_true',
+                        help="Avoid using CUDA when available")
+    parser.add_argument('--overwrite_output_dir', action='store_true',
+                        help="Overwrite the content of the output directory")
+    parser.add_argument('--seed', type=int, default=42,
+                        help="random seed for initialization")
+
+    parser.add_argument('--fp16', action='store_true',
+                        help="Whether to use 16-bit float precision instead of 32-bit")
+    parser.add_argument('--loss_scale', type=float, default=0,
+                        help="Loss scaling to improve fp16 numeric stability. Only used when fp16 set to True.\n"
+                             "0 (default value): dynamic loss scaling.\n"
+                             "Positive power of 2: static loss scaling value.\n")
+
+    parser.add_argument("--local_rank", type=int, default=-1,
+                        help="local_rank for distributed training on gpus")
+
+    parser.add_argument('--server_ip', type=str, default='', help="Can be used for distant debugging.")
+    parser.add_argument('--server_port', type=str, default='', help="Can be used for distant debugging.")
+    args = parser.parse_args()
+
+    # Setup distant debugging if needed
+    if args.server_ip and args.server_port:
+        # Distant debugging - see https://code.visualstudio.com/docs/python/debugging#_attach-to-a-local-script
+        import ptvsd
+        print("Waiting for debugger attach")
+        ptvsd.enable_attach(address=(args.server_ip, args.server_port), redirect_output=True)
+        ptvsd.wait_for_attach()
+
+    # Setup CUDA, GPU & distributed training
+    if args.local_rank == -1 or args.no_cuda:
+        device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
+        n_gpu = torch.cuda.device_count()
+    else:  # Initializes the distributed backend which will take care of sychronizing nodes/GPUs
+        torch.cuda.set_device(args.local_rank)
+        device = torch.device("cuda", args.local_rank)
+        torch.distributed.init_process_group(backend='nccl')
+        n_gpu = 1
+    args.device = device
+
+    # Setup logging
+    logging.basicConfig(level = logging.INFO if args.local_rank in [-1, 0] else logging.WARN)
+    logger.info("device: {} n_gpu: {}, distributed training: {}, 16-bits training: {}".format(
+                device, n_gpu, bool(args.local_rank != -1), args.fp16))
+
+    # Setup seeds
+    random.seed(args.seed)
+    np.random.seed(args.seed)
+    torch.manual_seed(args.seed)
+    if n_gpu > 0:
+        torch.cuda.manual_seed_all(args.seed)
+
+    # Safety checks and create output directory
+    if not args.do_train and not args.do_eval:
+        raise ValueError("At least one of `do_train` or `do_eval` must be True.")
+    if os.path.exists(args.output_dir) and os.listdir(args.output_dir) and args.do_train and not args.overwrite_output_dir:
+        raise ValueError("Output directory ({}) already exists and is not empty. Use --overwrite_output_dir to overcome.".format(args.output_dir))
+    if not os.path.exists(args.output_dir) and args.local_rank in [-1, 0]:
+        os.makedirs(args.output_dir)
+
+    # Prepare GLUE task
+    task_name = args.task_name.lower()
+    if task_name not in processors:
+        raise ValueError("Task not found: %s" % (task_name))
+    processor = processors[task_name]()
+    output_mode = output_modes[task_name]
+    label_list = processor.get_labels()
+    num_labels = len(label_list)
+
+    # Load pretrained model and tokenizer
+    if args.local_rank not in [-1, 0]:
+        # Make sure only the first process in distributed training will download model & vocab
+        torch.distributed.barrier()
+
+    tokenizer = BertTokenizer.from_pretrained(args.bert_model, do_lower_case=args.do_lower_case)
+    model = BertForSequenceClassification.from_pretrained(args.bert_model, num_labels=num_labels)
+
+    if args.local_rank == 0:
+        torch.distributed.barrier()
+
+    # Distributed, parrallel and fp16 model
+    if args.fp16:
+        model.half()
+    model.to(device)
+    if args.local_rank != -1:
+        model = torch.nn.parallel.DistributedDataParallel(model,
+                                                          device_ids=[args.local_rank],
+                                                          output_device=args.local_rank,
+                                                          find_unused_parameters=True)
+    elif n_gpu > 1:
+        model = torch.nn.DataParallel(model)
+
+    global_step = 0
+    tr_loss = 0
+    if args.do_train:
+        if args.local_rank in [-1, 0]:
+            tb_writer = SummaryWriter()
+
+        # Load and cache data
+        train_examples = processor.get_train_examples(args.data_dir)
+        cached_train_features_file = os.path.join(args.data_dir, 'train_{0}_{1}_{2}'.format(
+            list(filter(None, args.bert_model.split('/'))).pop(), str(args.max_seq_length), str(task_name)))
+        if os.path.exists(cached_train_features_file):
+            train_features = torch.load(cached_train_features_file)
+        else:
+            train_features = convert_examples_to_features(
+                train_examples, label_list, args.max_seq_length, tokenizer, output_mode)
+            if args.local_rank == -1 or torch.distributed.get_rank() == 0:
+                logger.info("  Saving train features into cached file %s", cached_train_features_file)
+                torch.save(train_features, cached_train_features_file)
+
+        # Convert in tensors and build dataloader
+        all_input_ids = torch.tensor([f.input_ids for f in train_features], dtype=torch.long)
+        all_input_mask = torch.tensor([f.input_mask for f in train_features], dtype=torch.long)
+        all_segment_ids = torch.tensor([f.segment_ids for f in train_features], dtype=torch.long)
+        if output_mode == "classification":
+            all_label_ids = torch.tensor([f.label_id for f in train_features], dtype=torch.long)
+        elif output_mode == "regression":
+            all_label_ids = torch.tensor([f.label_id for f in train_features], dtype=torch.float)
+
+        args.train_batch_size = args.train_batch_size // args.gradient_accumulation_steps
+
+        train_data = TensorDataset(all_input_ids, all_input_mask, all_segment_ids, all_label_ids)
+        train_sampler = RandomSampler(train_data) if args.local_rank == -1 else DistributedSampler(train_data)
+        train_dataloader = DataLoader(train_data, sampler=train_sampler, batch_size=args.train_batch_size)
+
+        num_train_optimization_steps = len(train_dataloader) // args.gradient_accumulation_steps * args.num_train_epochs
+
+        # Prepare optimizer
+        param_optimizer = list(model.named_parameters())
+        no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
+        optimizer_grouped_parameters = [
+            {'params': [p for n, p in param_optimizer if not any(nd in n for nd in no_decay)], 'weight_decay': 0.01},
+            {'params': [p for n, p in param_optimizer if any(nd in n for nd in no_decay)], 'weight_decay': 0.0}
+            ]
+        if args.fp16:
+            try:
+                from apex.optimizers import FP16_Optimizer, FusedAdam
+            except ImportError:
+                raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use fp16 training.")
+            optimizer = FusedAdam(optimizer_grouped_parameters,
+                                  lr=args.learning_rate,
+                                  bias_correction=False,
+                                  max_grad_norm=1.0)
+            if args.loss_scale == 0:
+                optimizer = FP16_Optimizer(optimizer, dynamic_loss_scale=True)
+            else:
+                optimizer = FP16_Optimizer(optimizer, static_loss_scale=args.loss_scale)
+            warmup_linear = WarmupLinearSchedule(warmup=args.warmup_proportion,
+                                                 t_total=num_train_optimization_steps)
+
+        else:
+            optimizer = BertAdam(optimizer_grouped_parameters,
+                                 lr=args.learning_rate,
+                                 warmup=args.warmup_proportion,
+                                 t_total=num_train_optimization_steps)
+
+        # Train!
+        logger.info("***** Running training *****")
+        logger.info("  Num examples = %d", len(train_examples))
+        logger.info("  Batch size = %d", args.train_batch_size)
+        logger.info("  Num steps = %d", num_train_optimization_steps)
+        model.train()
+        for _ in trange(int(args.num_train_epochs), desc="Epoch", disable=args.local_rank not in [-1, 0]):
+            for step, batch in enumerate(tqdm(train_dataloader, desc="Iteration", disable=args.local_rank not in [-1, 0])):
+                batch = tuple(t.to(device) for t in batch)
+                input_ids, input_mask, segment_ids, label_ids = batch
+
+                ouputs = model(input_ids, token_type_ids=segment_ids, attention_mask=input_mask, labels=label_ids)
+                loss = ouputs[0]
+
+                if n_gpu > 1:
+                    loss = loss.mean() # mean() to average on multi-gpu parallel training
+                if args.gradient_accumulation_steps > 1:
+                    loss = loss / args.gradient_accumulation_steps
+
+                if args.fp16:
+                    optimizer.backward(loss)
+                else:
+                    loss.backward()
+
+                tr_loss += loss.item()
+                if (step + 1) % args.gradient_accumulation_steps == 0:
+                    if args.fp16:
+                        # modify learning rate with special warm up BERT uses
+                        # if args.fp16 is False, BertAdam is used that handles this automatically
+                        lr_this_step = args.learning_rate * warmup_linear.get_lr(global_step, args.warmup_proportion)
+                        for param_group in optimizer.param_groups:
+                            param_group['lr'] = lr_this_step
+                    optimizer.step()
+                    optimizer.zero_grad()
+                    global_step += 1
+                    if args.local_rank in [-1, 0]:
+                        if not args.fp16:
+                            tb_writer.add_scalar('lr', optimizer.get_lr()[0], global_step)
+                        tb_writer.add_scalar('loss', loss.item(), global_step)
+
+    ### Saving best-practices: if you use defaults names for the model, you can reload it using from_pretrained()
+    ### Example:
+    if args.do_train and (args.local_rank == -1 or torch.distributed.get_rank() == 0):
+        # Save a trained model, configuration and tokenizer
+        model_to_save = model.module if hasattr(model, 'module') else model  # Only save the model it-self
+
+        # If we save using the predefined names, we can load using `from_pretrained`
+        output_model_file = os.path.join(args.output_dir, WEIGHTS_NAME)
+        output_config_file = os.path.join(args.output_dir, CONFIG_NAME)
+
+        torch.save(model_to_save.state_dict(), output_model_file)
+        model_to_save.config.to_json_file(output_config_file)
+        tokenizer.save_vocabulary(args.output_dir)
+
+        # Load a trained model and vocabulary that you have fine-tuned
+        model = BertForSequenceClassification.from_pretrained(args.output_dir)
+        tokenizer = BertTokenizer.from_pretrained(args.output_dir)
+
+        # Good practice: save your training arguments together with the trained model
+        output_args_file = os.path.join(args.output_dir, 'training_args.bin')
+        torch.save(args, output_args_file)
+    else:
+        model = BertForSequenceClassification.from_pretrained(args.bert_model)
+
+    model.to(device)
+
+    ### Evaluation
+    if args.do_eval and (args.local_rank == -1 or torch.distributed.get_rank() == 0):
+        eval_task_names = ("mnli", "mnli-mm") if task_name == "mnli" else (task_name,)
+        eval_outputs_dirs = (args.output_dir, args.output_dir + '-MM') if task_name == "mnli" else (args.output_dir,)
+        for eval_task, output_dir in zip(eval_task_names, eval_outputs_dirs):
+            if os.path.exists(output_dir) and os.listdir(output_dir) and args.do_train:
+                raise ValueError("Output directory ({}) already exists and is not empty.".format(args.output_dir))
+            if not os.path.exists(output_dir):
+                os.makedirs(output_dir)
+
+            # Load and cache data
+            processor = processors[eval_task]()
+            eval_examples = processor.get_dev_examples(args.data_dir)
+            cached_eval_features_file = os.path.join(args.data_dir, 'dev_{0}_{1}_{2}'.format(
+                list(filter(None, args.bert_model.split('/'))).pop(), str(args.max_seq_length), str(eval_task)))
+            if os.path.exists(cached_eval_features_file):
+                eval_features = torch.load(cached_eval_features_file)
+            else:
+                eval_features = convert_examples_to_features(
+                    eval_examples, label_list, args.max_seq_length, tokenizer, output_mode)
+                if args.local_rank == -1 or torch.distributed.get_rank() == 0:
+                    logger.info("  Saving eval features into cached file %s", cached_eval_features_file)
+                torch.save(eval_features, cached_eval_features_file)
+
+            # Convert in tensors and build dataloader
+            all_input_ids = torch.tensor([f.input_ids for f in eval_features], dtype=torch.long)
+            all_input_mask = torch.tensor([f.input_mask for f in eval_features], dtype=torch.long)
+            all_segment_ids = torch.tensor([f.segment_ids for f in eval_features], dtype=torch.long)
+            if output_mode == "classification":
+                all_label_ids = torch.tensor([f.label_id for f in eval_features], dtype=torch.long)
+            elif output_mode == "regression":
+                all_label_ids = torch.tensor([f.label_id for f in eval_features], dtype=torch.float)
+
+            eval_data = TensorDataset(all_input_ids, all_input_mask, all_segment_ids, all_label_ids)
+            # Note that DistributedSampler samples randomly
+            eval_sampler = SequentialSampler(eval_data) if args.local_rank == -1 else DistributedSampler(eval_data)
+            eval_dataloader = DataLoader(eval_data, sampler=eval_sampler, batch_size=args.eval_batch_size)
+
+            # Eval!
+            logger.info("***** Running evaluation *****")
+            logger.info("  Num examples = %d", len(eval_examples))
+            logger.info("  Batch size = %d", args.eval_batch_size)
+            model.eval()
+            eval_loss = 0
+            nb_eval_steps = 0
+            preds = None
+            out_label_ids = None
+            for batch in tqdm(eval_dataloader, desc="Evaluating"):
+                batch = tuple(t.to(device) for t in batch)
+                input_ids, input_mask, segment_ids, label_ids = batch
+
+                with torch.no_grad():
+                    outputs = model(input_ids,
+                                    token_type_ids=segment_ids,
+                                    attention_mask=input_mask,
+                                    labels=label_ids)
+                    tmp_eval_loss, logits = outputs[:2]
+
+                eval_loss += tmp_eval_loss.mean().item()
+                nb_eval_steps += 1
+                if preds is None:
+                    preds = logits.detach().cpu().numpy()
+                    out_label_ids = label_ids.detach().cpu().numpy()
+                else:
+                    preds = np.append(preds, logits.detach().cpu().numpy(), axis=0)
+                    out_label_ids = np.append(out_label_ids, label_ids.detach().cpu().numpy(), axis=0)
+
+            eval_loss = eval_loss / nb_eval_steps
+            if output_mode == "classification":
+                preds = np.argmax(preds, axis=1)
+            elif output_mode == "regression":
+                preds = np.squeeze(preds)
+            result = compute_metrics(eval_task, preds, out_label_ids)
+
+            loss = tr_loss/global_step if args.do_train else None
+
+            result['eval_loss'] = eval_loss
+            result['global_step'] = global_step
+            result['loss'] = loss
+
+            output_eval_file = os.path.join(output_dir, "eval_results.txt")
+            with open(output_eval_file, "w") as writer:
+                logger.info("***** Eval results *****")
+                for key in sorted(result.keys()):
+                    logger.info("  %s = %s", key, str(result[key]))
+                    writer.write("%s = %s\n" % (key, str(result[key])))
+
+
+if __name__ == "__main__":
+    main()
diff --git a/examples/utils_glue.py b/examples/utils_glue.py
index e3e4179fae..18e733567d 100644
--- a/examples/utils_glue.py
+++ b/examples/utils_glue.py
@@ -583,6 +583,7 @@ processors = {
 output_modes = {
     "cola": "classification",
     "mnli": "classification",
+    "mnli-mm": "classification",
     "mrpc": "classification",
     "sst-2": "classification",
     "sts-b": "regression",
diff --git a/pytorch_transformers/tokenization_utils.py b/pytorch_transformers/tokenization_utils.py
index c6f08c41ae..9004315657 100644
--- a/pytorch_transformers/tokenization_utils.py
+++ b/pytorch_transformers/tokenization_utils.py
@@ -110,6 +110,24 @@ class PreTrainedTokenizer(object):
 
         return tokenizer
 
+    def tokenize(self, text):
+        raise NotImplementedError
+
+    def convert_tokens_to_ids(self, tokens):
+        raise NotImplementedError
+
+    def convert_ids_to_tokens(self, ids):
+        raise NotImplementedError
+
+    def encode(self, text):
+        raise NotImplementedError
+
+    def decode(self, token_ids, *input, **kwargs):
+        raise NotImplementedError
+
+    def save_vocabulary(self, vocab_path):
+        raise NotImplementedError
+
 
 def clean_up_tokenization(out_string):
     out_string.replace(' .', '.').replace(' ?', '?').replace(' !', '!').replace(' ,', ','