May 4, 2019

2046 words 10 mins read



The Natural Language Decathlon: A Multitask Challenge for NLP

repo name salesforce/decaNLP
repo link
language Python
size (curr.) 844 kB
stars (curr.) 2036
created 2018-06-20
license BSD 3-Clause “New” or “Revised” License

decaNLP Logo

Build Status

The Natural Language Decathlon is a multitask challenge that spans ten tasks: question answering (SQuAD), machine translation (IWSLT), summarization (CNN/DM), natural language inference (MNLI), sentiment analysis (SST), semantic role labeling(QA‑SRL), zero-shot relation extraction (QA‑ZRE), goal-oriented dialogue (WOZ, semantic parsing (WikiSQL), and commonsense reasoning (MWSC). Each task is cast as question answering, which makes it possible to use our new Multitask Question Answering Network (MQAN). This model jointly learns all tasks in decaNLP without any task-specific modules or parameters in the multitask setting. For a more thorough introduction to decaNLP and the tasks, see the main website, our blog post, or the paper.

While the research direction associated with this repository focused on multitask learning, the framework itself is designed in a way that should make single-task training, transfer learning, and zero-shot evaluation simple. Similarly, the paper focused on multitask learning as a form of question answering, but this framework can be easily adapted for different approaches to single-task or multitask learning.


MQAN(Sampling+CoVe) 609.0 77.0 21.4 24.4 74.0 86.5 80.9 40.9 84.8 70.2 48.8
MQAN(QA‑first+CoVe) 599.9 75.5 18.9 24.4 73.6 86.4 80.8 37.4 85.8 68.5 48.8
MQAN(QA‑first) 590.5 74.4 18.6 24.3 71.5 87.4 78.4 37.6 84.8 64.8 48.7
S2S 513.6 47.5 14.2 25.7 60.9 85.9 68.7 28.5 84.0 45.8 52.4

Getting Started


The devices argument can be used to specify the devices for training. For CPU training, specify --devices -1; for GPU training, specify --devices DEVICEID. Note that Multi-GPU training is currently a WIP, so --device is sufficient for commands below. The default will be to train on GPU 0 as training on CPU will be quite time-consuming to train on all ten tasks in decaNLP.

If you want to use CPU, then remove the nvidia- and the cuda9_ prefixes from the default commands listed in sections below. This will allow you to use Docker without CUDA.

For example, if you have CUDA and all the necessary drivers and GPUs, you you can run a command inside the CUDA Docker image using:

nvidia-docker run -it --rm -v `pwd`:/decaNLP/ -u $(id -u):$(id -g) bmccann/decanlp:cuda9_torch041 bash -c "COMMAND --device 0"

If you want to run the same command without CUDA:

docker run -it --rm -v `pwd`:/decaNLP/ -u $(id -u):$(id -g) bmccann/decanlp:torch041 bash -c "COMMAND --device -1"

For those in the Docker know, you can look at the Dockerfiles used to build these two images in dockerfiles/.

PyTorch Version

The research associated with the original paper was done using Pytorch 0.3, but we have since migrated to 0.4. If you want to replicate results from the paper, then to be safe, you should use the code at a commit on or before 3c4f94b88768f4c3efc2fd4f015fed2f5453ebce. You should also replace toch041 with torch03 in the commands below to access a Docker image with the older version of PyTorch.


For example, to train a Multitask Question Answering Network (MQAN) on the Stanford Question Answering Dataset (SQuAD) on GPU 0:

nvidia-docker run -it --rm -v `pwd`:/decaNLP/ -u $(id -u):$(id -g) bmccann/decanlp:cuda9_torch041 bash -c "python /decaNLP/ --train_tasks squad --device 0"

To multitask with the fully joint, round-robin training described in the paper, you can add multiple tasks:

nvidia-docker run -it --rm -v `pwd`:/decaNLP/ -u $(id -u):$(id -g) bmccann/decanlp:cuda9_torch041 bash -c "python /decaNLP/ --train_tasks squad --train_iterations 1 --device 0"

To train on the entire Natural Language Decathlon:

nvidia-docker run -it --rm -v `pwd`:/decaNLP/ -u $(id -u):$(id -g) bmccann/decanlp:cuda9_torch041 bash -c "python /decaNLP/ --train_tasks squad cnn_dailymail sst srl zre woz.en wikisql schema --train_iterations 1 --device 0"

To pretrain on n_jump_start=1 tasks for jump_start=75000 iterations before switching to round-robin sampling of all tasks in the Natural Language Decathlon:

nvidia-docker run -it --rm -v `pwd`:/decaNLP/ -u $(id -u):$(id -g) bmccann/decanlp:cuda9_torch041 bash -c "python /decaNLP/ --n_jump_start 1 --jump_start 75000 --train_tasks squad cnn_dailymail sst srl zre woz.en wikisql schema --train_iterations 1 --device 0"

This jump starting (or pretraining) on a subset of tasks can be done for any set of tasks, not only the entirety of decaNLP.


If you would like to make use of tensorboard, you can add the --tensorboard flag to your training runs. This will log things in the format that Tensorboard expects.

To read those files and run the Tensorboard server, run (typically in a tmux pane or equivalent so that the process is not killed when you shut your laptop) the following command:

docker run -it --rm -p -v `pwd`:/decaNLP/ -u $(id -u):$(id -g) bmccann/decanlp:cuda9_torch041 bash -c "tensorboard --logdir /decaNLP/results"

If you are running the server on a remote machine, you can run the following on your local machine to forward to http://localhost:6006/:

ssh -4 -N -f -L 6006: YOUR_REMOTE_IP

If you are having trouble with the specified port on either machine, run lsof -if:6006 and kill the process if it is unnecessary. Otherwise, try changing the port numbers in the commands above. The first port number is the port the local machine tries to bind to, and and the second port is the one exposed by the remote machine (or docker container).

Notes on Training

  • On a single NVIDIA Volta GPU, the code should take about 3 days to complete 500k iterations. These should be sufficient to approximately reproduce the experiments in the paper. Training for about 7 days should be enough to fully replicate those scores, which should be only a few points higher than what is achieved by 500k iterations.
  • The model can be resumed using stored checkpoints using --load <PATH_TO_CHECKPOINT> and --resume. By default, models are stored every --save_every iterations in the results/ folder tree.
  • During training, validation can be slow! Especially when computing ROUGE scores. Use the --val_every flag to change the frequency of validation.
  • If you run out of GPU memory, reduce --train_batch_tokens and --val_batch_size.
  • If you run out of CPU memory, make sure that you are running the most recent version of the code that interns strings; if you are still running out of CPU memory, post an issue with the command you ran and your peak memory usage.
  • The first time you run, the code will download and cache all considered datasets. Please be advised that this might take a while, especially for some of the larger datasets.

Notes on Cached Data

  • In order to make data loading much quicker for repeated experiments, datasets are cached using code in text/torchtext/datasets/
  • If there is an update to this repository that touches any files in text/, then it might have changed the way a dataset is cached. If this is the case, then you’ll need to delete all relevant cached files or you will not see the changes.
  • Paths to cached files should be printed out when a dataset is loaded, either in training or in prediction. Search the text logged to stdout for Loading cached data from or Caching data to in order to locate the relevant path names for data caches.


You can evaluate a model for a specific task with EVALUATION_TYPE as validation or test:

nvidia-docker run -it --rm -v `pwd`:/decaNLP/ -u $(id -u):$(id -g) bmccann/decanlp:cuda9_torch041 bash -c "python /decaNLP/ --evaluate EVALUATION_TYPE --path PATH_TO_CHECKPOINT_DIRECTORY --device 0 --tasks squad"

or evaluate on the entire decathlon by removing any task specification:

nvidia-docker run -it --rm -v `pwd`:/decaNLP/ -u $(id -u):$(id -g) bmccann/decanlp:cuda9_torch041 bash -c "python /decaNLP/ --evaluate EVALUATION_TYPE --path PATH_TO_CHECKPOINT_DIRECTORY --device 0"

For test performance, please use the original SQuAD, MultiNLI, and WikiSQL evaluation systems. For WikiSQL, there is a detailed walk-through of how to get test numbers in the section of this document concerning pretrained models.

Pretrained Models

This model is the best MQAN trained on decaNLP so far. It was trained first on SQuAD and then on all of decaNLP. It uses CoVe as well. You can obtain this model and run it on the validation sets with the following.

tar -xvzf mqan_decanlp_better_sampling_cove_cpu.tgz
nvidia-docker run -it --rm -v `pwd`:/decaNLP/ -u $(id -u):$(id -g) bmccann/decanlp:cuda9_torch041 bash -c "python /decaNLP/ --evaluate validation --path /decaNLP/mqan_decanlp_better_sampling_cove_cpu/ --checkpoint_name iteration_560000.pth --device 0 --silent"

This model is the best MQAN trained on WikiSQL alone, which established a new state-of-the-art performance by several points on that task: 73.2 / 75.4 / 81.4 (ordered test logical form accuracy, unordered test logical form accuracy, test execution accuracy).

tar -xvzf mqan_wikisql_cpu.tar.gz
nvidia-docker run -it --rm -v `pwd`:/decaNLP/  bmccann/decanlp:cuda9_torch041 -c "python /decaNLP/ --evaluate validation --path /decaNLP/mqan_wikisql_cpu --checkpoint_name iteration_57000.pth --device 0 --tasks wikisql"
nvidia-docker run -it --rm -v `pwd`:/decaNLP/ -u $(id -u):$(id -g) bmccann/decanlp:cuda9_torch041 bash -c "python /decaNLP/ --evaluate test --path /decaNLP/mqan_wikisql_cpu --checkpoint_name iteration_57000.pth --device 0 --tasks wikisql"
docker run -it --rm -v `pwd`:/decaNLP/ -u $(id -u):$(id -g) bmccann/decanlp:cuda9_torch041 bash -c "python /decaNLP/ /decaNLP/.data/ /decaNLP/mqan_wikisql_cpu/iteration_57000/validation/wikisql.txt /decaNLP/mqan_wikisql_cpu/iteration_57000/validation/wikisql.ids.txt /decaNLP/mqan_wikisql_cpu/iteration_57000/validation/wikisql_logical_forms.jsonl valid"
docker run -it --rm -v `pwd`:/decaNLP/ -u $(id -u):$(id -g) bmccann/decanlp:cuda9_torch041 bash -c "python /decaNLP/ /decaNLP/.data/ /decaNLP/mqan_wikisql_cpu/iteration_57000/test/wikisql.txt /decaNLP/mqan_wikisql_cpu/iteration_57000/test/wikisql.ids.txt /decaNLP/mqan_wikisql_cpu/iteration_57000/test/wikisql_logical_forms.jsonl test"
git clone for ssh
docker run -it --rm -v `pwd`:/decaNLP/ -u $(id -u):$(id -g) bmccann/decanlp:cuda9_torch041 bash -c "python /decaNLP/WikiSQL/ /decaNLP/.data/wikisql/data/dev.jsonl /decaNLP/.data/wikisql/data/dev.db /decaNLP/mqan_wikisql_cpu/iteration_57000/validation/wikisql_logical_forms.jsonl" # assumes that you have data stored in .data
docker run -it --rm -v `pwd`:/decaNLP/ -u $(id -u):$(id -g) bmccann/decanlp:cuda9_torch041 bash -c "python /decaNLP/WikiSQL/ /decaNLP/.data/wikisql/data/test.jsonl /decaNLP/.data/wikisql/data/test.db /decaNLP/mqan_wikisql_cpu/iteration_57000/test/wikisql_logical_forms.jsonl" # assumes that you have data stored in .data

You can similarly follow the instructions above for downloading, decompressing, and loading in pretrained models for other indivual tasks (single-task models):


Inference on a Custom Dataset

Using a pretrained model or a model you have trained yourself, you can run on new, custom datasets easily by following the instructions below. In this example, we use the checkpoint for the best MQAN trained on the entirety of decaNLP (see the section on Pretrained Models to see how to get this checkpoint) to run on my_custom_dataset.

mkdir -p .data/my_custom_dataset/
touch .data/my_custom_dataset/val.jsonl
echo '{"context": "The answer is answer.", "question": "What is the answer?", "answer": "answer"}' >> .data/my_custom_dataset/val.jsonl 
# TODO add your own examples line by line to val.jsonl in the form of a JSON dictionary, as demonstrated above.
# Make sure to delete the first line if you don't want the demonstrated example.
nvidia-docker run -it --rm -v `pwd`:/decaNLP/ -u $(id -u):$(id -g) bmccann/decanlp:cuda9_torch041 bash -c "python /decaNLP/ --evaluate valid --path /decaNLP/mqan_decanlp_qa_first_cpu --checkpoint_name iteration_1140000.pth --tasks my_custom_dataset"

You should get output that ends with something like this:

**  /decaNLP/mqan_decanlp_qa_first_cpu/iteration_1140000/valid/my_custom_dataset.txt  already exists -- this is where predictions are stored **
**  /decaNLP/mqan_decanlp_qa_first_cpu/modeltion_1140000/valid/  already exists -- this is where ground truth answers are stored **
**  /decaNLP/mqan_decanlp_qa_first_cpu/modeltion_1140000/valid/my_custom_dataset.results.txt  already exists -- this is where metrics are stored **

{'em': 0.0, 'nf1': 100.0, 'nem': 100.0}
Prediction: the answer
Answer: answer

From this output, you can see where predictions are stored along with ground truth outputs and metrics. If you want to rerun using this model checkpoint on this particular dataset, you’ll need to pass the --overwrite_predictions argument to If you do not want predictions and answers printed to stdout, then pass the --silent argument to

The metrics dictionary should have printed something like {'em': 0.0, 'nf1': 100.0, 'nem': 100.0}. Here em stands for exact match. This is the percentage of predictions that had every token match the ground truth answer exactly. The normalized version, nem, lowercases and strips punctuation – all of our models are trained on lowercased data, so nem is a more accurate representation of performance than em for our models. For tasks that are typically treated as classification problems, these exact match scores should correspond to accuracy. nf1 is a normalized (lowercased; punctuation stripped) F1 score over the predicted and ground truth sequences. If you would like to add additional metrics that are already implemented you can try adding --bleu (the typical metric for machine translation) and --rouge (the typical metric for summarization). Other metrics can be implemented following the patterns in


If you use this in your work, please cite The Natural Language Decathlon: Multitask Learning as Question Answering.

  title={The Natural Language Decathlon: Multitask Learning as Question Answering},
  author={Bryan McCann and Nitish Shirish Keskar and Caiming Xiong and Richard Socher},
  journal={arXiv preprint arXiv:1806.08730},


Contact: and

comments powered by Disqus