Transformers >= 4.36.1
This model relies on a custom modeling file, you need to add trust_remote_code=True
See #13467

LSG ArXiv paper.
Github/conversion script is available at this link.

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM, pipeline

tokenizer = AutoTokenizer.from_pretrained("ccdv/lsg-bart-base-4096-wcep", trust_remote_code=True)
model = AutoModelForSeq2SeqLM.from_pretrained("ccdv/lsg-bart-base-4096-wcep", trust_remote_code=True)

text = "Replace by what you want."
pipe = pipeline("text2text-generation", model=model, tokenizer=tokenizer, device=0)
generated_text = pipe(text, truncation=True, max_length=64, no_repeat_ngram_size=7)

ccdv/lsg-bart-base-4096-wcep

This model is a fine-tuned version of ccdv/lsg-bart-base-4096 on the ccdv/WCEP-10 roberta dataset.
It achieves the following results on the test set:

Length	Sparse Type	Block Size	Sparsity	Connexions	R1	R2	RL	RLsum
4096	Local	256	0	768	46.02	24.23	37.38	38.72
4096	Local	128	0	384	45.43	23.86	36.94	38.30
4096	Pooling	128	4	644	45.36	23.61	36.75	38.06
4096	Stride	128	4	644	45.87	24.31	37.41	38.70
4096	Block Stride	128	4	644	45.78	24.16	37.20	38.48
4096	Norm	128	4	644	45.34	23.39	36.47	37.78
4096	LSH	128	4	644	45.15	23.53	36.74	38.02

With smaller block size (lower ressources):

Length	Sparse Type	Block Size	Sparsity	Connexions	R1	R2	RL	RLsum
4096	Local	64	0	192	44.48	22.98	36.20	37.52
4096	Local	32	0	96	43.60	22.17	35.61	36.66
4096	Pooling	32	4	160	43.91	22.41	35.80	36.92
4096	Stride	32	4	160	44.62	23.11	36.32	37.53
4096	Block Stride	32	4	160	44.47	23.02	36.28	37.46
4096	Norm	32	4	160	44.45	23.03	36.10	37.33
4096	LSH	32	4	160	43.87	22.50	35.75	36.93

Model description

The model relies on Local-Sparse-Global attention to handle long sequences:

The model has about ~145 millions parameters (6 encoder layers - 6 decoder layers).
The model is warm started from BART-base, converted to handle long sequences (encoder only) and fine tuned.

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 8e-05
• train_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 4
• total_train_batch_size: 32
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_ratio: 0.1
• num_epochs: 10.0

Generate hyperparameters

The following hyperparameters were used during generation:

• dataset_name: ccdv/WCEP-10
• dataset_config_name: roberta
• eval_batch_size: 8
• eval_samples: 1022
• early_stopping: True
• ignore_pad_token_for_loss: True
• length_penalty: 2.0
• max_length: 64
• min_length: 0
• num_beams: 5
• no_repeat_ngram_size: None
• seed: 123

Framework versions

• Transformers 4.18.0
• Pytorch 1.10.1+cu102
• Datasets 2.1.0
• Tokenizers 0.11.6