BiGS is the first model to achieve BERT-level transfer learning on the GLUE benchmark with subquadratic complexity in length (or without attention). It is also the first bidirectional state space model in NLP.
Torch Masked Language Model
import torch
from BiGS.modeling_bigs import BiGSForMaskedLM
model = BiGSForMaskedLM.from_pretrained('JunxiongWang/BiGS_128')Torch Sequence Classification Model
import torch
from BiGS.modeling_bigs import BiGSForSequenceClassification
model = BiGSForSequenceClassification.from_pretrained('JunxiongWang/BiGS_128')For GLUE task, please see GLUE_torch.md and GLUE_torch_freeze.md. If you don't want to use MNLI checkpoints to finetune MRPC, RTE, STS-B, please run GLUE_torch_freeze.md. Notice that torch version has slight worse results compared with Jax version.
Paper |
This repository contains BiGS's jax model definitions, pretrained models weights, training and fintuning code for our paper exploring using state space models for pretraining. You can find more details in our paper.
Pretraining Without Attention
Junxiong Wang, Jing Nathan Yan, Albert Gu, Alexander M.Rush
Cornell University, Cornell Tech, DeepMind
Transformers have been essential to pretraining success in NLP. While other architectures have been used, downstream accuracy is either significantly worse, or requires attention layers to match standard benchmarks such as GLUE. This work explores pretraining without attention by using recent advances in sequence routing based on state-space models (SSMs). Our proposed model, Bidirectional Gated SSM (BiGS), combines SSM layers with a multiplicative gating architecture that has been effective in simplified sequence modeling architectures. The model learns static layers that do not consider pair-wise interactions. Even so, BiGS is able to match BERT pretraining accuracy on GLUE and can be extended to long-form pretraining of 4096 tokens without approximation. Analysis shows that while the models have similar accuracy, the approach has significantly different inductive biases than BERT in terms of interactions and syntactic representations.
This repo contains:
- 🪐 JAX implementation of BiGS and its variants,
- 🛸 Pre-trained BiGS Models of various lengths,
- 💥 Training scripts to train BiGS from scratch,
- 💫 Fine-tuning scripts for GLUE tasks
You can run our models on both GPUs and TPUs.
For TPUs,
pip install -r requirements-tpu.txt
For GPUs,
pip install -r requirements-gpu.txt
| Sentence Length | Trained Tokens | Link |
|---|---|---|
| 128 | ~11B | BiGS-11B-128 |
| 128 | ~29B | BiGS-29B-128 |
| 128 | ~97B | BiGS-97B-128 |
| 512 | ~108B | BiGS-108B-512 |
| 1024 | ~110B | BiGS-110B-1024 |
| 4096 | ~110B | BiGS-110B-4096 |
| Sentence Length | Trained Tokens | Model |
|---|---|---|
| 128 | ~11B | BiGS-11B-128MNLI |
| 128 | ~29B | BiGS-29B-128MNLI |
| 128 | ~97B | BiGS-97B-128MNLI |
| 512 | ~108B | BiGS-108B-512MNLI |
import jax
from jax import numpy as jnp
from transformers import BertTokenizer
from BiGS.modeling_flax_bigs import FlaxBiGSForMaskedLM
tokenizer = BertTokenizer.from_pretrained('bert-large-uncased')
model = FlaxBiGSForMaskedLM.from_pretrained('JunxiongWang/BiGS_128')
text = "The goal of life is [MASK]."
encoded_input = tokenizer(text, return_tensors='np', padding='max_length', max_length=128)
output = model(**encoded_input)
tokenizer.convert_ids_to_tokens(jnp.flip(jnp.argsort(jax.nn.softmax(output.logits[encoded_input['input_ids']==103]))[0])[:10])
# output: ['happiness', 'love', 'peace', 'perfection', 'life', 'enlightenment', 'god', 'survival', 'freedom', 'good']
jnp.flip(jnp.sort(jax.nn.softmax(output.logits[encoded_input['input_ids']==103]))[0])[:10]
# probability: [0.16052087, 0.04306792, 0.03651363, 0.03468223, 0.02927081, 0.02549769, 0.02385132, 0.02261189, 0.01672831, 0.01619471]
text = "Paris is the [MASK] of France."
encoded_input = tokenizer(text, return_tensors='np', padding='max_length', max_length=128)
output = model(**encoded_input)
tokenizer.convert_ids_to_tokens(jnp.flip(jnp.argsort(jax.nn.softmax(output.logits[encoded_input['input_ids']==103]))[0])[:8])
# output: ['capital', 'centre', 'center', 'city', 'capitol', 'prefecture', 'headquarters', 'president', 'metropolis', 'heart']
jnp.flip(jnp.sort(jax.nn.softmax(output.logits[encoded_input['input_ids']==103]))[0])[:10]
# probability: [0.9981787 , 0.00034076, 0.00026992, 0.00026926, 0.00017787, 0.00004816, 0.00004256, 0.00003716, 0.00003634, 0.00002893]from BiGS.modeling_flax_bigs import FlaxBiGSForSequenceClassification
model = FlaxBiGSForSequenceClassification.from_pretrained('JunxiongWang/BiGS_512')from BiGS.modeling_flax_bigs import FlaxBiGSForQuestionAnswering
model = FlaxBiGSForQuestionAnswering.from_pretrained('JunxiongWang/BiGS_512')from BiGS.modeling_flax_bigs import FlaxBiGSForMultipleChoice
model = FlaxBiGSForMultipleChoice.from_pretrained('JunxiongWang/BiGS_512')See pretrain.md
See GLUE.md and GLUE_freeze.md. If you don't want to use MNLI checkpoints to finetune MRPC, RTE, STS-B, please run GLUE_freeze.md.
If you find this project inspiring or use our repository, kindly please cite
@article{wang2022pretraining,
title={Pretraining Without Attention},
author={Wang, Junxiong and Yan, Jing Nathan and Gu, Albert and Rush, Alexander M},
journal={arXiv preprint arXiv:2212.10544},
year={2022}
}