flan-t5-base-samsum / README.md
philschmid's picture
philschmid HF staff
update model card README.md
8df1497
---
license: apache-2.0
tags:
- generated_from_trainer
datasets:
- samsum
metrics:
- rouge
model-index:
- name: flan-t5-base-samsum
results:
- task:
name: Sequence-to-sequence Language Modeling
type: text2text-generation
dataset:
name: samsum
type: samsum
config: samsum
split: train
args: samsum
metrics:
- name: Rouge1
type: rouge
value: 47.2358
---
<!-- This model card has been generated automatically according to the information the Trainer had access to. You
should probably proofread and complete it, then remove this comment. -->
# flan-t5-base-samsum
This model is a fine-tuned version of [google/flan-t5-base](https://huggingface.co/google/flan-t5-base) on the samsum dataset.
It achieves the following results on the evaluation set:
- Loss: 1.3716
- Rouge1: 47.2358
- Rouge2: 23.5135
- Rougel: 39.6266
- Rougelsum: 43.3458
- Gen Len: 17.3907
## 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 | Validation Loss | Rouge1 | Rouge2 | Rougel | Rougelsum | Gen Len |
|:-------------:|:-----:|:----:|:---------------:|:-------:|:-------:|:-------:|:---------:|:-------:|
| 1.4379 | 1.0 | 1842 | 1.3805 | 47.1075 | 23.531 | 39.6919 | 43.549 | 17.1197 |
| 1.3559 | 2.0 | 3684 | 1.3716 | 47.2358 | 23.5135 | 39.6266 | 43.3458 | 17.3907 |
| 1.2783 | 3.0 | 5526 | 1.3721 | 47.4581 | 23.7339 | 39.7726 | 43.4568 | 17.1832 |
| 1.2378 | 4.0 | 7368 | 1.3757 | 47.8557 | 24.0593 | 40.2324 | 44.0085 | 17.3053 |
| 1.1983 | 5.0 | 9210 | 1.3751 | 47.8156 | 24.0038 | 40.2169 | 43.8918 | 17.3040 |
### Framework versions
- Transformers 4.25.1
- Pytorch 1.12.1+cu113
- Datasets 2.8.0
- Tokenizers 0.12.1