Fine-tuned on a synthetic dataset of curated long-context text and GPT-3.5-turbo-1106 summaries spanning multiple domains + "random" long-context examples from pretraining datasets

Note: this model has not been fine-tuned on any other summarization datasets, just the synthsumm data

Try it: gradio demo | free HF inference api via requests| .md with example outputs (gauntlet)

Usage

It's recommended to use this model with beam search decoding. If interested, you can also use the textsum util repo to have most of this abstracted out for you:

pip install -U textsum

python

from textsum.summarize import Summarizer

model_name = "pszemraj/long-t5-tglobal-base-synthsumm_direct"
summarizer = Summarizer(model_name) # GPU auto-detected
text = "put the text you don't want to read here"
summary = summarizer.summarize_string(text)
print(summary)

Details

This model is a fine-tuned version of google/long-t5-tglobal-base on the None dataset. It achieves the following results on the evaluation set:

Loss: 1.4378
Rouge1: 48.0918
Rouge2: 21.2531
Rougel: 34.4307
Rougelsum: 43.0271
Gen Len: 84.5231

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 0.0003
train_batch_size: 1
eval_batch_size: 1
seed: 26605
gradient_accumulation_steps: 8
total_train_batch_size: 8
optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
lr_scheduler_type: inverse_sqrt
lr_scheduler_warmup_ratio: 0.03
num_epochs: 2.0

Training results

Training Loss	Epoch	Step	Validation Loss	Rouge1	Rouge2	Rougel	Rougelsum	Gen Len
1.9183	0.38	125	1.5762	38.7221	15.0873	28.3123	34.9655	129.2154
1.8815	0.77	250	1.5230	44.3531	17.9384	31.7417	39.5563	87.3538
1.7264	1.15	375	1.4735	45.7781	20.102	33.329	41.4737	101.9231
1.8545	1.54	500	1.4505	47.0134	20.6159	33.6118	41.6579	88.2308
1.7444	1.92	625	1.4378	48.0918	21.2531	34.4307	43.0271	84.5231

Framework versions

Transformers 4.36.0.dev0
Pytorch 2.1.0
Datasets 2.15.0
Tokenizers 0.15.0