---
license: cc-by-sa-4.0
pipeline_tag: fill-mask
arxiv: 2210.05529
language: en
thumbnail: https://github.com/coastalcph/hierarchical-transformers/raw/main/data/figures/hat_encoder.png
tags:
- long-documents
datasets:
- c4
model-index:
- name: kiddothe2b/hierarchical-transformer-base-4096
  results: []
---

# Hierarchical Attention Transformer (HAT) / hierarchical-transformer-base-4096

## Model description

This is a Hierarchical Attention Transformer (HAT) model as presented in [An Exploration of Hierarchical Attention Transformers for Efficient Long Document Classification (Chalkidis et al., 2022)](https://arxiv.org/abs/2210.05529). 

The model has been warm-started re-using the weights of RoBERTa (Liu et al., 2019), and continued pre-trained for MLM in long sequences following the paradigm of Longformer released by Beltagy et al. (2020). It supports sequences of length up to 4,096.

HAT uses hierarchical attention, which is a combination of segment-wise and cross-segment attention operations. You can think of segments as paragraphs or sentences.

## Intended uses & limitations

You can use the raw model for masked language modeling, but it's mostly intended to be fine-tuned on a downstream task.
See the [model hub](https://huggingface.co/models?filter=hierarchical-transformer) to look for other versions of HAT or fine-tuned versions on a task that interests you.

Note that this model is primarily aimed at being fine-tuned on tasks that use the whole document to make decisions, such as document classification, sequential sentence classification, or question answering.

## How to use

You can use this model directly for masked language modeling:

```python
from transformers import AutoTokenizer, AutoModelForForMaskedLM
tokenizer = AutoTokenizer.from_pretrained("kiddothe2b/hierarchical-transformer-base-4096", trust_remote_code=True)
mlm_model = AutoModelForMaskedLM("kiddothe2b/hierarchical-transformer-base-4096", trust_remote_code=True)
```

You can also fine-tune it for SequenceClassification, SequentialSentenceClassification, and MultipleChoice down-stream tasks:

```python
from transformers import AutoTokenizer, AutoModelForSequenceClassification
tokenizer = AutoTokenizer.from_pretrained("kiddothe2b/hierarchical-transformer-base-4096", trust_remote_code=True)
doc_classifier = AutoModelForSequenceClassification.from_pretrained("kiddothe2b/hierarchical-transformer-base-4096", trust_remote_code=True)
```

## Limitations and bias

The training data used for this model contains a lot of unfiltered content from the internet, which is far from
neutral. Therefore, the model can have biased predictions.


## Training procedure

### Training and evaluation data

The model has been warm-started from [roberta-base](https://huggingface.co/roberta-base) checkpoint and has been continued pre-trained for additional 50k steps in long sequences (> 1024 subwords) of [C4](https://huggingface.co/datasets/c4) (Raffel et al., 2020).


### Training hyperparameters

The following hyperparameters were used during training:
- learning_rate: 0.0001
- train_batch_size: 2
- eval_batch_size: 2
- seed: 42
- distributed_type: tpu
- num_devices: 8
- gradient_accumulation_steps: 8
- total_train_batch_size: 128
- total_eval_batch_size: 16
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- lr_scheduler_warmup_ratio: 0.1
- training_steps: 50000

### Training results

| Training Loss | Epoch | Step  | Validation Loss |
|:-------------:|:-----:|:-----:|:---------------:|
| 1.7437        | 0.2   | 10000 | 1.6370          |
| 1.6994        | 0.4   | 20000 | 1.6054          |
| 1.6726        | 0.6   | 30000 | 1.5718          |
| 1.644         | 0.8   | 40000 | 1.5526          |
| 1.6299        | 1.0   | 50000 | 1.5368          |


### Framework versions

- Transformers 4.19.0.dev0
- Pytorch 1.11.0+cu102
- Datasets 2.0.0
- Tokenizers 0.11.6


## Citing

If you use HAT in your research, please cite:

[An Exploration of Hierarchical Attention Transformers for Efficient Long Document Classification](https://arxiv.org/abs/2210.05529). Ilias Chalkidis, Xiang Dai, Manos Fergadiotis, Prodromos Malakasiotis, and Desmond Elliott. 2022. arXiv:2210.05529 (Preprint).

```
@misc{chalkidis-etal-2022-hat,
  url = {https://arxiv.org/abs/2210.05529},
  author = {Chalkidis, Ilias and Dai, Xiang and Fergadiotis, Manos and Malakasiotis, Prodromos and Elliott, Desmond},
  title = {An Exploration of Hierarchical Attention Transformers for Efficient Long Document Classification},
  publisher = {arXiv},
  year = {2022},
}
```