---
tags:
- summarization
- summary
- booksum
- long-document
- long-form
- tglobal-xl
- XL
license:
- apache-2.0
- bsd-3-clause
datasets:
- kmfoda/booksum
metrics:
- rouge
inference: false
model-index:
- name: pszemraj/long-t5-tglobal-xl-16384-book-summary
  results:
  - task:
      type: summarization
      name: Summarization
    dataset:
      name: multi_news
      type: multi_news
      config: default
      split: test
    metrics:
    - name: ROUGE-1
      type: rouge
      value: 36.2043
      verified: true
    - name: ROUGE-2
      type: rouge
      value: 8.424
      verified: true
    - name: ROUGE-L
      type: rouge
      value: 17.3721
      verified: true
    - name: ROUGE-LSUM
      type: rouge
      value: 32.3994
      verified: true
    - name: loss
      type: loss
      value: 2.0843334197998047
      verified: true
    - name: gen_len
      type: gen_len
      value: 248.3572
      verified: true
---

# long-t5-tglobal-xl + BookSum

Summarize long text and get a SparkNotes-esque summary of arbitrary topics!
- Generalizes reasonably well to academic & narrative text.
- This is the XL checkpoint, which **from a human-evaluation perspective, [produces even better summaries](https://long-t5-xl-book-summary-examples.netlify.app/)**.

A simple example/use case with [the base model](https://huggingface.co/pszemraj/long-t5-tglobal-base-16384-book-summary) on ASR is [here](https://longt5-booksum-example.netlify.app/).

## Cheeky Proof-of-Concept

A summary of the [infamous navy seals copypasta](https://knowyourmeme.com/memes/navy-seal-copypasta):

> In this chapter, the monster explains how he intends to exact revenge on "the little b****" who insulted him. He tells the kiddo that he is a highly trained and experienced killer who will use his arsenal of weapons--including his access to the internet--to exact justice on the little brat.

While a somewhat crude example, try running this copypasta through other summarization models to see the difference in comprehension (_despite it not even being a "long" text!_)

---

## Description

A fine-tuned version of [google/long-t5-tglobal-xl](https://huggingface.co/google/long-t5-tglobal-xl) on the `kmfoda/booksum` dataset.

Read the paper by Guo et al. here: [LongT5: Efficient Text-To-Text Transformer for Long Sequences](https://arxiv.org/pdf/2112.07916.pdf) 

## How-To in Python

> 🚧 `LLM.int8()` appears to be compatible with summarization and does not degrade the quality of the outputs; this is a crucial enabler for using this model on standard GPUs. A PR for this is in-progress [here](https://github.com/huggingface/transformers/pull/20341), and this model card will be updated with instructions once done :) 🚧

Install/update transformers `pip install -U transformers`

Summarize text with pipeline:

```python
import torch
from transformers import pipeline

summarizer = pipeline(
    "summarization",
    "pszemraj/long-t5-tglobal-xl-16384-book-summary",
    device=0 if torch.cuda.is_available() else -1,
)
long_text = "Here is a lot of text I don't want to read. Replace me"

result = summarizer(long_text)
print(result[0]["summary_text"])
```

Pass [other parameters related to beam search textgen](https://huggingface.co/blog/how-to-generate) when calling `summarizer` to get even higher quality results.

---

## About

### Intended uses & limitations

While this model seems to improve upon factual consistency, **do not take summaries to be foolproof and check things that seem odd**.

Specifically: negation statements (i.e., model says: _This thing does not have [ATTRIBUTE]_ where instead it should have said _This thing has a lot of [ATTRIBUTE]_).
- I'm sure someone will write a paper on this eventually (if there isn't one already), but you can usually fact-check this by comparing a specific claim to what the surrounding sentences imply.

### Training and evaluation data

`kmfoda/booksum` dataset on HuggingFace - read [the original paper here](https://arxiv.org/abs/2105.08209). 

- **Initial fine-tuning** only used input text with 12288 tokens input or less and 1024 tokens output or less (_i.e. rows with longer were dropped before training_) for memory reasons. Per brief analysis, summaries in the 12288-16384 range in this dataset are in the **small** minority
  - In addition, this initial training combined the training and validation sets and trained on these in aggregate to increase the functional dataset size. **Therefore, take the validation set results with a grain of salt; primary metrics should be (always) the test set.**
- **final phases of fine-tuning** used the standard conventions of 16384 input/1024 output keeping everything (truncating longer sequences). This did not appear to change the loss/performance much.

### Eval results

Official results with the [model evaluator](https://huggingface.co/spaces/autoevaluate/model-evaluator) will be computed and posted here.

**Please read the note above as due to training methods, validation set performance looks better than the test set results will be**. The model achieves the following results on the evaluation set:
- eval_loss: 1.2756
- eval_rouge1: 41.8013
- eval_rouge2: 12.0895
- eval_rougeL: 21.6007
- eval_rougeLsum: 39.5382
- eval_gen_len: 387.2945
- eval_runtime: 13908.4995
- eval_samples_per_second: 0.107
- eval_steps_per_second: 0.027

```
***** predict/test metrics (initial) *****                                                               
  predict_gen_len            =   506.4368                                                 
  predict_loss               =      2.028                                                 
  predict_rouge1             =    36.8815                                                 
  predict_rouge2             =     8.0625                                                 
  predict_rougeL             =    17.6161                                                 
  predict_rougeLsum          =    34.9068                                                 
  predict_runtime            = 2:04:14.37                                                 
  predict_samples            =       1431                                                 
  predict_samples_per_second =      0.192                                                 
  predict_steps_per_second   =      0.048
```
\* evaluating big model not as easy as it seems. Doing a bit more investigating

---

## FAQ

### How can I run inference with this on CPU?

lol

### How to run inference over a very long (30k+ tokens) document in batches?

See `summarize.py` in [the code for my hf space Document Summarization](https://huggingface.co/spaces/pszemraj/document-summarization/blob/main/summarize.py) :)

You can also use the same code to split a document into batches of 4096, etc., and run over those with the model. This is useful in situations where CUDA memory is limited.

### How to fine-tune further?

See [train with a script](https://huggingface.co/docs/transformers/run_scripts) and [the summarization scripts](https://github.com/huggingface/transformers/tree/main/examples/pytorch/summarization)


---

## Training procedure

### Updates

Updates to this model/model card will be posted here as relevant. The model seems fairly converged; if updates/improvements are possible using the `BookSum` dataset, this repo will be updated.

### Training hyperparameters

The following hyperparameters were used during training:
- learning_rate: 0.0006
- train_batch_size: 1
- eval_batch_size: 1
- seed: 10350
- distributed_type: multi-GPU
- num_devices: 4
- gradient_accumulation_steps: 32
- total_train_batch_size: 128
- total_eval_batch_size: 4
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: constant
- num_epochs: 1.0

\*_Prior training sessions used roughly similar parameters (learning rates were higher); multiple sessions were required as this takes eons to train._

### Framework versions

- Transformers 4.25.0.dev0
- Pytorch 1.13.0+cu117
- Datasets 2.6.1
- Tokenizers 0.13.1

---