flan-t5-base-tldr-100k

This model is a fine-tuned version of google/flan-t5-base on the first 100.000 samples of the Reddit TL;DR dataset. It achieves the following results on the evaluation set:

• Loss: 2.7323
• Rouge1: 17.0772
• Rouge2: 4.4204
• Rougel: 14.549
• Rougelsum: 15.0148
• Gen Len: 16.0925

Model description

More information needed

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: 5e-05
• train_batch_size: 8
• eval_batch_size: 8
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• num_epochs: 5

Training results

Training Loss	Epoch	Step	Gen Len	Validation Loss	Rouge1	Rouge2	Rougel	Rougelsum
3.1467	1.0	11250	16.4911	2.9531	15.909	3.8156	13.4884	13.9568
3.0673	2.0	22500	16.4639	2.9318	16.4972	3.9757	13.9517	14.4336
2.9952	3.0	33750	16.2585	2.9245	16.5997	4.1068	14.0299	14.5147
2.9524	4.0	45000	2.7323	17.0772	4.4204	14.549	15.0148	16.0925
2.9223	5.0	56250	2.7328	17.1468	4.4384	14.5798	15.0572	16.2163

Framework versions

• Transformers 4.27.4
• Pytorch 2.0.0+cu117
• Datasets 2.11.0
• Tokenizers 0.13.2