flan-t5-base-finetuned-length_control_token

This model is a fine-tuned version of google/flan-t5-base on an unknown dataset. It achieves the following results on the evaluation set:

• Loss: 1.0276
• Sacrebleu: 16.2445

Model description

This model was trained on a dataset called PWKP-GPT3-LENGTH-CONTROL-40BUCKETS. The dataset contains 30k instances taken from PWKP, then processed through GPT3 to obtain simplifications. The 30k instances come from: 10k which were supposed to generate very long simplifications, 10k which were supposed to generate very short simplifications, and 10k without specifying the simplicity level. The model does not sucessfuly work on these buckets. There exists another dataset, the PWKP-GPT3-LENGTH-CONTROL-4BUCKETS, but it was never trained on something. Those buckets are also rather unbalanced.

The idea comes from Controllable Sentence Simplification Louis Martin, https://arxiv.org/pdf/1910.02677.pdf

It was fine-tuned on the FLAN-T5-base model.

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 5.6e-05
• train_batch_size: 16
• eval_batch_size: 16
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• num_epochs: 6

Training results

Training Loss	Epoch	Step	Validation Loss	Sacrebleu
1.3257	1.0	1782	1.0906	15.4208
1.1718	2.0	3564	1.0648	15.5358
1.0972	3.0	5346	1.0484	15.8113
1.0472	4.0	7128	1.0394	16.0159
1.0092	5.0	8910	1.0305	16.1341
0.9858	6.0	10692	1.0276	16.2445

Framework versions

• Transformers 4.26.1
• Pytorch 1.13.1+cu117
• Datasets 2.10.1
• Tokenizers 0.13.2