This is a ported version of fairseq-based [wmt16 transformer](https://github.com/jungokasai/deep-shallow/) for {src_lang}-{tgt_lang}.
For more details, please, see [Deep Encoder, Shallow Decoder: Reevaluating the Speed-Quality Tradeoff in Machine Translation](https://arxiv.org/abs/2006.10369).
Pretrained weights were left identical to the original model released by allenai. For more details, please, see the [paper](https://arxiv.org/abs/2006.10369).
The score is slightly below the score reported in the paper, as the researchers don't use `sacrebleu` and measure the score on tokenized outputs. `transformers` score was measured using `sacrebleu` on detokenized outputs.
This is a ported version of fairseq-based wmt19 transformer created by [jungokasai]](https://github.com/jungokasai/) @ allenai for {src_lang}-{tgt_lang}.
<!-- This file has been auto-generated by src/transformers/convert_fsmt_original_pytorch_checkpoint_to_pytorch.py - DO NOT EDIT or your changes will be lost -->
This is a ported version of [fairseq wmt19 transformer](https://github.com/pytorch/fairseq/blob/master/examples/wmt19/README.md) for {src_lang}-{tgt_lang}.
For more details, please see, [Facebook FAIR's WMT19 News Translation Task Submission](https://arxiv.org/abs/1907.06616).
The abbreviation FSMT stands for FairSeqMachineTranslation
- The original (and this ported model) doesn't seem to handle well inputs with repeated sub-phrases, [content gets truncated](https://discuss.huggingface.co/t/issues-with-translating-inputs-containing-repeated-phrases/981)
## Training data
Pretrained weights were left identical to the original model released by fairseq. For more details, please, see the [paper](https://arxiv.org/abs/1907.06616).
## Eval results
pair | fairseq | transformers
-------|---------|----------
{pair} | {scores[pair][0]} | {scores[pair][1]}
The score is slightly below the score reported by `fairseq`, since `transformers`` currently doesn't support:
- model ensemble, therefore the best performing checkpoint was ported (``model4.pt``).
