ul2-large-dutch-simplification-mai-2023

This model is intended to simplify Dutch sentences.

This model is a fine-tuned version of yhavinga/ul2-large-dutch on the BramVanroy/chatgpt-dutch-simplification dataset.

The model was created in light of the master thesis of Charlotte Van de Velde in the Master of Science in Artificial Intelligence (MAI) at KU Leuven in 2023. Charlotte is supervised by Vincent Vandeghinste and Bram Vanroy. Dataset creation by Charlotte, model training by Bram.

Quick links

• Repository: includes training code and model creation log
• Dataset: BramVanroy/chatgpt-dutch-simplification
• Parent model: this model was finetuned on yhavinga/ul2-large-dutch
• Demo: shows the "base" model in action (don't rely on the "Hosted inference API" widget on this page, it does not work very well)

Intended uses & limitations, and dataset

The model is intended for sentence-level simplification of Dutch. It might extend to document-level simplification but most of the dataset is limited to sentences so document-level performance is not guaranteed.

The dataset has been generated automatically (cf. dataset description) and has not been manually verified. On top of that, this model has been fine-tuned and we did not scrutinize the parent model or its training data. Output of the current model is therefore subject to unexpected results (as most if not all neural networks).

Because the dataset was generated with ChatGPT, this model cannot be used for commercial purposes.

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0002927210895006501
• train_batch_size: 32
• optimizer: Adafactor
• num_epochs: 27

These hyperarameters were found through Bayesian hyperparameter search with wandb. This is described in the repository.

Training results

eval results are on the evaluation set, predict results are on the test set. These were achieved with beam search (num_beams=3).

{
    "eval_gen_len": 21.404761904761905,
    "eval_loss": 3.0882697105407715,
    "eval_rouge1": 41.3871,
    "eval_rouge2": 19.6751,
    "eval_rougeL": 36.0469,
    "eval_rougeLsum": 36.1178,
    "eval_sari": 54.3588,
  
    "predict_gen_len": 22.1484375,
    "predict_loss": 2.7822625637054443,
    "predict_rouge1": 43.8191,
    "predict_rouge2": 21.7783,
    "predict_rougeL": 39.3657,
    "predict_rougeLsum": 39.3751,
    "predict_sari": 52.3752
}

Note: the model seems to underperform compared to the base variant of the model, achieving only similar results with a much larger size. The reason for this may be found in the hyperparameters, where this large model may have benefitted from a smaller learning rate in the optimisation space. In the hyperparameter search, the learning rate spectrum was set to 1e-03 to 1e-04 but this might be too large for this model and size.

Framework versions

• Transformers 4.29.2
• Pytorch 2.0.1+cu117
• Datasets 2.12.0
• Tokenizers 0.13.3