Skip to content


Repository files navigation


YANMTT is short for Yet Another Neural Machine Translation Toolkit. For a backstory how I ended up creating this toolkit scroll to the bottom of this README. Although the name says that it is yet another toolkit, it was written with the purpose of better understanding of the flow of training, starting from data pre-processing, sharding, batching, distributed training and decoding. There is a significant emphashis on multilingualism and on cross-lingual learning.

List of features:

  1. Basic NMT pre-training, fine-tuning, decoding, visualization
    • Distributed, mixed precision, multilingual training.
    • Denoising pre-training in mBART, mT5 or UL2 style.
    • Fine-tuning your own or official BART-like models like BART, mBART, IndicBART.
    • Joint supervised and unsupervised training using monolingual and parallel corpora.
    • Sentence representation, attention extraction, and scoring translations.
  2. User Interface
    • GUI to demo and debug models.
    • Select any official huggingface (mBART, IndicBART) or custom model and run it on any of the supported languages.
    • Visualize attention weights at each layer for each head using bertviz.
    • Visualize encoder representations of a set of sentences using tensorflow projector.
  3. Advanced features
    • Multi-billion parameter models can be trained using FSDP.
    • Mixtures-of-experts layers.
    • Tempered softmax training.
    • Softmax calibration.
    • Entropy maximization training.
    • Multi-layer softmax training.
    • 8-bit optimizers for training large models.
    • Various weight initialization strategies.
    • Various positional embedding strategies like sinusoidal, learned, RoPE, AliBi, NoPE.
  4. Light-weight fine-tuning
    • Adaptor (Houlsby, Bapna, Mixtures of Adapters) and prompt tuning.
    • Hypercomplex, IA-3 light-weight adaptors.
    • Eliminate components or layers prior to decoding or fine-tuning.
    • Fine-grained control over what parameters to fine-tune.
  5. Model compression
    • Training compact models from scratch via recurrently stacked layers (ALBART).
    • Distillation of pre-trained and fine-tuned models.
  6. Simultaneous NMT
    • Simulated Wait-k NMT where we train and decode wait-K models or decode full-sentence models using wait-k.
  7. Multi-source and Document NMT
    • Vanilla multi-source with two input sentences belonging to different languages.
    • Document level NMT where one input is the current sentence and the other one is the context.
    • Various multi-source fusion strategies.
    • Can be combined with wait-k NMT.

How to install: Follow installation_instructions.txt

Installing the GUI:

  1. Follow the file in the interface folder.
  2. The GUI does not explicitly depend on YANMTT.

Scripts and their functionality:

  1. and These scripts govern the creation of a unigram SPM or BPE tokenizer. The shell script creates the subword segmenter using sentencepiece which can make both SPM and BPE models. All you need is a monolingual corpus for the languages you are interested in. The python script wraps this around an AlbertTokenizer (for SPM) or MBartTokenizer (for BPE), adds special user defined tokens and saves a configuration file for use in the future via an AutoTokenizer.
    Usage: see examples/

  2. This is used to pre-train a model. At the very least you need a monolingual corpus for the languages you are interested in and a tokenizer trained for those languages. This script can also be used to do denoising style training jointly with regular NMT training although the NMT training is rather basic because there is no evaluation during training. If you want to do advanced NMT training then you should use the "" script. Ultimately, you should not use the outcome of this script to perform final translations. Additional advanced usages involve: simulated wait-k simultaneous NMT, knowledge distillation, fine-tuning pre-existing MBART models with fine-grained control over what should be initialized, frozen or tuned, etc. Read the code and the command line arguments for a better understanding of the advanced features.
    Usage: see examples/
    Note 1: If M is your model name then a folder "M_deploy" is created which you can directly use with AutoTokenizer and AutoModel.
    Note 2: If you plan to use this "M_deploy" model with the GUI then remember to use the --supported_languages flag.

  3. This is used to either train a NMT model from scratch or fine-tune a pre-existing MBART or a NMT model created via YANMTT. At the very least you need a parallel corpus (preferrably split into train, dev and test sets although we can make do with only a train set) for the language pairs you are interested in. There are several advanced features such as: simulated wait-k simultaneous NMT, knowledge distillation, fine-grained control over what should be initialized, frozen or tuned, document NMT, multi-source NMT, adaptor tuning, prompt tuning, mixtures of experts layers, multilingual NMT training.
    Usage: see examples/ Note: The notes applying to the "" script also apply to this script.

  4. This is used to decode sentences using a trained model. Additionally you can do translation pair scoring, forced decoding, forced alignment (experimental), encoder/decoder representation extraction and alignment visualization.
    Usage: see examples/

  5. This contains all housekeeping functions such as corpora splitting, batch generation, loss computation etc. Do take a look at all the methods since you may need to modify them.

  6. You can average the specified checkpoints using arithmetic averaging.
    Usage: see examples/

  7. This is used to temporarily occupy a gpu in case you use a shared GPU environment. Run this in the background before launching the training processes so that while the training scripts are busy doing preprocessing like sharding or model loading, the GPU you aim for is not occupied by someone else. Usage will be shown in the example scripts for training.


  1. Whenever running the example usage scripts simply run them as examples/ from the root directory of the toolkit
  2. The data under examples/data is taken from and is released the ALT Parallel Corpus as a Creative Commons Attribution 4.0 International (CC BY 4.0)

License and copyright:

  1. MIT licence for code that I wrote.
  2. Apache licence for modifications or additions to the huggingface code in the transformers folder.

Copyright 2021 National Institute of Information and Communication Technology (Raj Dabre)

Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions: The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software. THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.

Contact me (Raj Dabre) at or for general queries. For queries about the user interface, please contact Diptesh Kanojia at or Chinmay Sawant at

Backstory: Why I made this toolkit
Despite the fact that I enjoy coding, I never really pushed myself throughout my Masters and Ph.D. towards writing a self contained toolkit. I had always known that coding is an important part of research and although I had made plenty of meaningful changes to several code bases, I never felt like I owned any of those changes.

Fast forward to 2020 where I wanted to play with MBART/BART/MASS. It would have been easy to use fairseq or tensor2tensor but then again the feeling of lack of ownership would remain. Huggingface provides a lot of implementations but (at the time) had no actual script to easily do MBART pre-training. All I had was this single comment and this guide for distributed training using pytorch (thanks yangkky).

After a bit of hesitation I decided to get my hands dirty and make a quick notebook for MBART pretraining. That snowballed into me writing my own pipeline for data sharding, preprocessing and training. Since I was at it I wrote a pipeline for tine tuning. Why not go further and write a pipeline for decoding and analysis? Fine-grained control over fine-tuning? Distillation? Multi-source NMT? Document NMT? Simultaneous Wait-K NMT? 3 months later I ended up with this toolkit which I wanted to share with everyone. Since I have worked in low-resource MT and efficent MT this toolkit will mostly contain implementations that somehow involve transfer learning, compression/distillation, simultaneous NMT.

I am pretty sure its not as fast or perfect like the ones written by the awesome people at GAFA but I will be more than happy if a few people use my toolkit.