Sequence to Sequence Training and Evaluation

This directory contains examples for finetuning and evaluating transformers on summarization and translation tasks. Please tag @patil-suraj with any issues/unexpected behaviors, or send a PR! For deprecated bertabs instructions, see bertabs/README.md. For the old finetune_trainer.py and related utils, see examples/legacy/seq2seq.

Supported Architectures

• BartForConditionalGeneration
• FSMTForConditionalGeneration (translation only)
• MBartForConditionalGeneration
• MarianMTModel
• PegasusForConditionalGeneration
• T5ForConditionalGeneration

run_summarization.py and run_translation.py are lightweight examples of how to download and preprocess a dataset from the 🤗 Datasets library or use your own files (jsonlines or csv), then fine-tune one of the architectures above on it.

For custom datasets in jsonlines format please see: https://huggingface.co/docs/datasets/loading_datasets.html#json-files and you also will find examples of these below.

Summarization

Here is an example on a summarization task:

python examples/seq2seq/run_summarization.py \
    --model_name_or_path t5-small \
    --do_train \
    --do_eval \
    --dataset_name cnn_dailymail \
    --dataset_config "3.0.0" \
    --source_prefix "summarize: " \
    --output_dir /tmp/tst-summarization \
    --per_device_train_batch_size=4 \
    --per_device_eval_batch_size=4 \
    --overwrite_output_dir \
    --predict_with_generate

Only T5 models t5-small, t5-base, t5-large, t5-3b and t5-11b must use an additional argument: --source_prefix "summarize: ".

We used CNN/DailyMail dataset in this example as t5-small was trained on it and one can get good scores even when pre-training with a very small sample.

Extreme Summarization (XSum) Dataset is another commonly used dataset for the task of summarization. To use it replace --dataset_name cnn_dailymail --dataset_config "3.0.0" with --dataset_name xsum.

And here is how you would use it on your own files, after adjusting the values for the arguments --train_file, --validation_file, --text_column and --summary_column to match your setup:

python examples/seq2seq/run_summarization.py \
    --model_name_or_path t5-small \
    --do_train \
    --do_eval \
    --train_file path_to_csv_or_jsonlines_file \
    --validation_file path_to_csv_or_jsonlines_file \
    --source_prefix "summarize: " \
    --output_dir /tmp/tst-summarization \
    --overwrite_output_dir \
    --per_device_train_batch_size=4 \
    --per_device_eval_batch_size=4 \
    --predict_with_generate

The task of summarization supports custom CSV and JSONLINES formats.

Custom CSV Files

If it's a csv file the training and validation files should have a column for the inputs texts and a column for the summaries.

If the csv file has just two columns as in the following example:

text,summary
"I'm sitting here in a boring room. It's just another rainy Sunday afternoon. I'm wasting my time I got nothing to do. I'm hanging around I'm waiting for you. But nothing ever happens. And I wonder","I'm sitting in a room where I'm waiting for something to happen"
"I see trees so green, red roses too. I see them bloom for me and you. And I think to myself what a wonderful world. I see skies so blue and clouds so white. The bright blessed day, the dark sacred night. And I think to myself what a wonderful world.","I'm a gardener and I'm a big fan of flowers."
"Christmas time is here. Happiness and cheer. Fun for all that children call. Their favorite time of the year. Snowflakes in the air. Carols everywhere. Olden times and ancient rhymes. Of love and dreams to share","It's that time of year again."

The first column is assumed to be for text and the second is for summary.

If the csv file has multiple columns, you can then specify the names of the columns to use:

    --text_column text_column_name \
    --summary_column summary_column_name \

For example if the columns were:

id,date,text,summary

and you wanted to select only text and summary, then you'd pass these additional arguments:

    --text_column text \
    --summary_column summary \

Custom JSONLINES Files

The second supported format is jsonlines. Here is an example of a jsonlines custom data file.

{"text": "I'm sitting here in a boring room. It's just another rainy Sunday afternoon. I'm wasting my time I got nothing to do. I'm hanging around I'm waiting for you. But nothing ever happens. And I wonder", "summary": "I'm sitting in a room where I'm waiting for something to happen"}
{"text": "I see trees so green, red roses too. I see them bloom for me and you. And I think to myself what a wonderful world. I see skies so blue and clouds so white. The bright blessed day, the dark sacred night. And I think to myself what a wonderful world.", "summary": "I'm a gardener and I'm a big fan of flowers."}
{"text": "Christmas time is here. Happiness and cheer. Fun for all that children call. Their favorite time of the year. Snowflakes in the air. Carols everywhere. Olden times and ancient rhymes. Of love and dreams to share", "summary": "It's that time of year again."}

Same as with the CSV files, by default the first value will be used as the text record and the second as the summary record. Therefore you can use any key names for the entries, in this example text and summary were used.

And as with the CSV files, you can specify which values to select from the file, by explicitly specifying the corresponding key names. In our example this again would be:

    --text_column text \
    --summary_column summary \

Translation

Here is an example of a translation fine-tuning with a MarianMT model:

python examples/seq2seq/run_translation.py \
    --model_name_or_path Helsinki-NLP/opus-mt-en-ro \
    --do_train \
    --do_eval \
    --source_lang en \
    --target_lang ro \
    --dataset_name wmt16 \
    --dataset_config_name ro-en \
    --output_dir /tmp/tst-translation \
    --per_device_train_batch_size=4 \
    --per_device_eval_batch_size=4 \
    --overwrite_output_dir \
    --predict_with_generate

MBart and some T5 models require special handling.

T5 models t5-small, t5-base, t5-large, t5-3b and t5-11b must use an additional argument: --source_prefix "translate {source_lang} to {target_lang}". For example:

python examples/seq2seq/run_translation.py \
    --model_name_or_path t5-small \
    --do_train \
    --do_eval \
    --source_lang en \
    --target_lang ro \
    --source_prefix "translate English to Romanian: " \
    --dataset_name wmt16 \
    --dataset_config_name ro-en \
    --output_dir /tmp/tst-translation \
    --per_device_train_batch_size=4 \
    --per_device_eval_batch_size=4 \
    --overwrite_output_dir \
    --predict_with_generate

If you get a terrible BLEU score, make sure that you didn't forget to use the --source_prefix argument.

For the aforementioned group of T5 models it's important to remember that if you switch to a different language pair, make sure to adjust the source and target values in all 3 language-specific command line argument: --source_lang, --target_lang and --source_prefix.

MBart models require a different format for --source_lang and --target_lang values, e.g. instead of en it expects en_XX, for ro it expects ro_RO. The full MBart specification for language codes can be found here. For example:

python examples/seq2seq/run_translation.py \
    --model_name_or_path facebook/mbart-large-en-ro  \
    --do_train \
    --do_eval \
    --dataset_name wmt16 \
    --dataset_config_name ro-en \
    --source_lang en_XX \
    --target_lang ro_RO \
    --output_dir /tmp/tst-translation \
    --per_device_train_batch_size=4 \
    --per_device_eval_batch_size=4 \
    --overwrite_output_dir \
    --predict_with_generate

And here is how you would use the translation finetuning on your own files, after adjusting the values for the arguments --train_file, --validation_file to match your setup:

python examples/seq2seq/run_translation.py \
    --model_name_or_path t5-small \
    --do_train \
    --do_eval \
    --source_lang en \
    --target_lang ro \
    --source_prefix "translate English to Romanian: " \
    --dataset_name wmt16 \
    --dataset_config_name ro-en \
    --train_file path_to_jsonlines_file \
    --validation_file path_to_jsonlines_file \
    --output_dir /tmp/tst-translation \
    --per_device_train_batch_size=4 \
    --per_device_eval_batch_size=4 \
    --overwrite_output_dir \
    --predict_with_generate

The task of translation supports only custom JSONLINES files, with each line being a dictionary with a key "translation" and its value another dictionary whose keys is the language pair. For example:

{ "translation": { "en": "Others have dismissed him as a joke.", "ro": "Alții l-au numit o glumă." } }
{ "translation": { "en": "And some are holding out for an implosion.", "ro": "Iar alții așteaptă implozia." } }

Here the languages are Romanian (ro) and English (en).

If you want to use a pre-processed dataset that leads to high BLEU scores, but for the en-de language pair, you can use --dataset_name stas/wmt14-en-de-pre-processed, as following:

python examples/seq2seq/run_translation.py \
    --model_name_or_path t5-small \
    --do_train \
    --do_eval \
    --source_lang en \
    --target_lang de \
    --source_prefix "translate English to German: " \
    --dataset_name stas/wmt14-en-de-pre-processed \
    --output_dir /tmp/tst-translation \
    --per_device_train_batch_size=4 \
    --per_device_eval_batch_size=4 \
    --overwrite_output_dir \
    --predict_with_generate