metadata
license: apache-2.0
base_model: t5-small
tags:
- generated_from_trainer
datasets:
- xsum
metrics:
- rouge
model-index:
- name: fastSUMMARIZER-t5-small-finetuned-on-xsum
results:
- task:
name: Sequence-to-sequence Language Modeling
type: text2text-generation
dataset:
name: xsum
type: xsum
config: default
split: validation
args: default
metrics:
- name: Rouge1
type: rouge
value: 31.3222
pipeline_tag: summarization
widget:
- text: >-
There will soon be flying taxis. Many of us grew up watching science
fiction movies with these. The Japanese airline ANA and a U.S. tech
start-up called Joby Aviation will fly air taxis at the 2025 World Expo in
Osaka. They are currently building the taxis. They will need to follow air
traffic rules. They will also need to train flying taxi pilots. The
five-seat, all-electric taxi will take off and land vertically. It will
fly as far as 241 kilometers and have a top speed of 321kph. Joby said the
taxis are environmentally friendly. People can reduce their carbon
footprint. It said Japan was a great place to test the taxis because 92
per cent of the population live in towns and cities. The president of ANA
said the airline has 70 years of safe and reliable flights. He said it was
good that customers have 'the option to travel rapidly, and sustainably,
from an international airport to a downtown location'.
- text: >-
Everybody knows that eating carrots is good for our eyesight. A new study
suggests that grapes are also good for our eyes. Researchers from the
National University of Singapore have found that eating just a few grapes
a day can improve our vision. This is especially so for people who are
older. Dr Eun Kim, the lead researcher, said: 'Our study is the first to
show that grape consumption beneficially impacts eye health in humans,
which is very exciting, especially with a growing, ageing population.' Dr
Kim added that, 'grapes are an easily accessible fruit that studies have
shown can have a beneficial impact' on our eyesight. This is good news for
people who don't really like carrots. The study is published in the
journal 'Food & Function'. Thirty-four adults took part in a series of
experiments over 16 weeks. Half of the participants ate one-and-a-half
cups of grapes per day; the other half ate a placebo snack. Dr Kim did not
tell the participants or the researchers whether she was testing the
grapes or the snack. She thought that not revealing this information would
give better test results. She found that people who ate the grapes had
improved muscle strength around the retina. The retina passes information
about light to the brain via electrical signals. It protects the eyes from
damaging blue light. A lot of blue light comes from computer and
smartphone screens, and from LED lights.
t5-small-finetuned-summarization-xsum
This model is a fine-tuned version of t5-small on the xsum dataset. It is very fast and light. The model summarizes a whole text in just <1s, making it very efficient for low resource usage.
Model Demo:
https://huggingface.co/spaces/Rahmat82/RHM-text-summarizer-light
It achieves the following results on the evaluation set:
- Loss: 2.2425
- Rouge1: 31.3222
- Rouge2: 10.0614
- Rougel: 25.0513
- Rougelsum: 25.0446
- Gen Len: 18.8044
Model description
This model is light and performs very fast. No matter on GPU or CPU, it always summarizes your text in <1s. If you use optimum, it may get even faster.
Click the following link to open the model's demo:
https://huggingface.co/spaces/Rahmat82/RHM-text-summarizer-light
Use the model:
from transformers import AutoModelForSeq2SeqLM, AutoTokenizer, pipeline
model_id = "Rahmat82/t5-small-finetuned-summarization-xsum"
model = AutoModelForSeq2SeqLM.from_pretrained(model_id)
tokenizer = AutoTokenizer.from_pretrained(model_id, use_fast=True)
summarizer = pipeline("summarization",model = model, tokenizer=tokenizer)
text_to_summarize = """
The koala is regarded as the epitome of cuddliness. However, animal lovers
will be saddened to hear that this lovable marsupial has been moved to the
endangered species list. The Australian Koala Foundation estimates there are
somewhere between 43,000-100,000 koalas left in the wild. Their numbers have
been dwindling rapidly due to disease, loss of habitat, bushfires, being hit
by cars, and other threats. Stuart Blanch from the World Wildlife Fund in
Australia said: "Koalas have gone from no listing to vulnerable to endangered
within a decade. That is a shockingly fast decline." He added that koalas risk
"sliding toward extinction"
"""
print(summarizer(text_to_summarize)[0]["summary_text"])
Use model with optimum/onnxruntime - super fast:
#!pip install -q transformers accelerate optimum onnxruntime onnx
from transformers import AutoTokenizer
from optimum.onnxruntime import ORTModelForSeq2SeqLM
from optimum.pipelines import pipeline
import accelerate
model_name = "Rahmat82/t5-small-finetuned-summarization-xsum"
model = ORTModelForSeq2SeqLM.from_pretrained(model_name, export=True)
tokenizer = AutoTokenizer.from_pretrained(model_name, use_fast=True)
summarizer = pipeline("summarization", model=model, tokenizer=tokenizer,
device_map="auto", batch_size=12)
text_to_summarize = """
The koala is regarded as the epitome of cuddliness. However, animal lovers
will be saddened to hear that this lovable marsupial has been moved to the
endangered species list. The Australian Koala Foundation estimates there are
somewhere between 43,000-100,000 koalas left in the wild. Their numbers have
been dwindling rapidly due to disease, loss of habitat, bushfires, being hit
by cars, and other threats. Stuart Blanch from the World Wildlife Fund in
Australia said: "Koalas have gone from no listing to vulnerable to endangered
within a decade. That is a shockingly fast decline." He added that koalas risk
"sliding toward extinction"
"""
print(summarizer(text_to_summarize)[0]["summary_text"])
Training hyperparameters
The following hyperparameters were used during training:
- learning_rate: 0.0002
- train_batch_size: 28
- eval_batch_size: 28
- seed: 42
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- num_epochs: 2
- mixed_precision_training: Native AMP
Training results
Training Loss | Epoch | Step | Validation Loss | Rouge1 | Rouge2 | Rougel | Rougelsum | Gen Len |
---|---|---|---|---|---|---|---|---|
2.5078 | 1.0 | 7288 | 2.2860 | 30.9087 | 9.7673 | 24.6951 | 24.6927 | 18.7973 |
2.4245 | 2.0 | 14576 | 2.2425 | 31.3222 | 10.0614 | 25.0513 | 25.0446 | 18.8044 |
Framework versions
- Transformers 4.37.0
- Pytorch 2.1.2
- Datasets 2.1.0
- Tokenizers 0.15.1