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

Try it: gradio demo | example outputs .md (gauntlet) | code for free HF inference api

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/pegasus-x-large-book_synthsumm"
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 pszemraj/pegasus-x-large-book-summary on the None dataset. It achieves the following results on the evaluation set:

Loss: 1.5481
Rouge1: 48.141
Rouge2: 19.1137
Rougel: 33.647
Rougelsum: 42.1211
Gen Len: 73.9846

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 0.0003
train_batch_size: 1
eval_batch_size: 1
seed: 5309
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
1.7369	0.38	125	1.7140	43.0265	15.8613	30.5774	38.2507	77.0462
1.7736	0.77	250	1.6361	43.0209	15.2384	29.7678	37.4955	67.6
1.4251	1.15	375	1.5931	46.2138	17.5559	33.0091	41.0385	74.1077
1.2706	1.54	500	1.5635	44.6382	16.5917	30.7551	39.8466	71.7231
1.4844	1.92	625	1.5481	48.141	19.1137	33.647	42.1211	73.9846

Framework versions

Transformers 4.36.0.dev0
Pytorch 2.1.0
Datasets 2.15.0
Tokenizers 0.15.0