# MosaicBERT-Base model

MosaicBERT-Base is a new BERT architecture and training recipe optimized for fast pretraining. MosaicBERT trains faster and achieves higher pretraining and finetuning accuracy when benchmarked against Hugging Face's bert-base-uncased.

March 2023

## How to use

from transformers import AutoModelForMaskedLM


The tokenizer for this model is simply the Hugging Face bert-base-uncased tokenizer.

from transformers import BertTokenizer
tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')


To use this model directly for masked language modeling, use pipeline:

from transformers import AutoModelForMaskedLM, BertTokenizer, pipeline

tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')

classifier("I [MASK] to the store yesterday.")


To continue MLM pretraining, follow the MLM pre-training section of the mosaicml/examples/bert repo.

To fine-tune this model for classification, follow the Single-task fine-tuning section of the mosaicml/examples/bert repo.

### Remote Code

This model requires that trust_remote_code=True be passed to the from_pretrained method. This is because we train using FlashAttention (Dao et al. 2022), which is not part of the transformers library and depends on Triton and some custom PyTorch code. Since this involves executing arbitrary code, you should consider passing a git revision argument that specifies the exact commit of the code, for example:

mlm = AutoModelForMaskedLM.from_pretrained(
'mosaicml/mosaic-bert-base',
trust_remote_code=True,
revision='24512df',
)


However, if there are updates to this model or code and you specify a revision, you will need to manually check for them and update the commit hash accordingly.

## Model description

In order to build MosaicBERT, we adopted architectural choices from the recent transformer literature. These include FlashAttention (Dao et al. 2022), ALiBi (Press et al. 2021), and Gated Linear Units (Shazeer 2020). In addition, we remove padding inside the transformer block, and apply LayerNorm with low precision.

### Modifications to the Attention Mechanism

1. FlashAttention: Attention layers are core components of the transformer architecture. The recently proposed FlashAttention layer reduces the number of read/write operations between the GPU HBM (high bandwidth memory, i.e. long-term memory) and the GPU SRAM (i.e. short-term memory) [Dao et al. 2022]. We used the FlashAttention module built by hazy research with OpenAI’s triton library.

2. Attention with Linear Biases (ALiBi): In most BERT models, the positions of tokens in a sequence are encoded with a position embedding layer; this embedding allows subsequent layers to keep track of the order of tokens in a sequence. ALiBi eliminates position embeddings and instead conveys this information using a bias matrix in the attention operation. It modifies the attention mechanism such that nearby tokens strongly attend to one another [Press et al. 2021]. In addition to improving the performance of the final model, ALiBi helps the model to handle sequences longer than it saw during training. Details on our ALiBi implementation can be found in the mosaicml/examples repo here.

3. Unpadding: Standard NLP practice is to combine text sequences of different lengths into a batch, and pad the sequences with empty tokens so that all sequence lengths are the same. During training, however, this can lead to many superfluous operations on those padding tokens. In MosaicBERT, we take a different approach: we concatenate all the examples in a minibatch into a single sequence of batch size 1. Results from NVIDIA and others have shown that this approach leads to speed improvements during training, since operations are not performed on padding tokens (see for example Zeng et al. 2022). Details on our “unpadding” implementation can be found in the mosaicml/examples repo here.

4. Low Precision LayerNorm: this small tweak forces LayerNorm modules to run in float16 or bfloat16 precision instead of float32, improving utilization. Our implementation can be found in the mosaicml/examples repo here.

### Modifications to the Feedforward Layers

1. Gated Linear Units (GLU): We used Gated Linear Units for the feedforward sublayer of a transformer. GLUs were first proposed in 2016 [Dauphin et al. 2016], and incorporate an extra learnable matrix that “gates” the outputs of the feedforward layer. More recent work has shown that GLUs can improve performance quality in transformers [Shazeer, 2020, Narang et al. 2021]. We used the GeLU (Gaussian-error Linear Unit) activation function with GLU, which is sometimes referred to as GeGLU. The GeLU activation function is a smooth, fully differentiable approximation to ReLU; we found that this led to a nominal improvement over ReLU. More details on our implementation of GLU can be found here. The extra gating matrix in a GLU model potentially adds additional parameters to a model; we chose to augment our BERT-Base model with additional parameters due to GLU modules as it leads to a Pareto improvement across all timescales (which is not true of all larger models such as BERT-Large). While BERT-Base has 110 million parameters, MosaicBERT-Base has 137 million parameters. Note that MosaicBERT-Base trains faster than BERT-Base despite having more parameters.

## Training data

MosaicBERT is pretrained using a standard Masked Language Modeling (MLM) objective: the model is given a sequence of text with some tokens hidden, and it has to predict these masked tokens. MosaicBERT is trained on the English “Colossal, Cleaned, Common Crawl” C4 dataset, which contains roughly 365 million curated text documents scraped from the internet (equivalent to 156 billion tokens). We used this more modern dataset in place of traditional BERT pretraining corpora like English Wikipedia and BooksCorpus.

## Pretraining Optimizations

Many of these pretraining optimizations below were informed by our BERT results for the MLPerf v2.1 speed benchmark.

1. MosaicML Streaming Dataset: As part of our efficiency pipeline, we converted the C4 dataset to MosaicML’s StreamingDataset format and used this for both MosaicBERT-Base and the baseline BERT-Base. For all BERT-Base models, we chose the training duration to be 286,720,000 samples of sequence length 128; this covers 78.6% of C4.

2. Higher Masking Ratio for the Masked Language Modeling Objective: We used the standard Masked Language Modeling (MLM) pretraining objective. While the original BERT paper also included a Next Sentence Prediction (NSP) task in the pretraining objective, subsequent papers have shown this to be unnecessary Liu et al. 2019. However, we found that a 30% masking ratio led to slight accuracy improvements in both pretraining MLM and downstream GLUE performance. We therefore included this simple change as part of our MosaicBERT training recipe. Recent studies have also found that this simple change can lead to downstream improvements Wettig et al. 2022.

3. Bfloat16 Precision: We use bf16 (bfloat16) mixed precision training for all the models, where a matrix multiplication layer uses bf16 for the multiplication and 32-bit IEEE floating point for gradient accumulation. We found this to be more stable than using float16 mixed precision.

4. Vocab Size as a Multiple of 64: We increased the vocab size to be a multiple of 8 as well as 64 (i.e. from 30,522 to 30,528). This small constraint is something of a magic trick among ML practitioners, and leads to a throughput speedup.

5. Hyperparameters: For all models, we use Decoupled AdamW with Beta_1=0.9 and Beta_2=0.98, and a weight decay value of 1.0e-5. The learning rate schedule begins with a warmup to a maximum learning rate of 5.0e-4 followed by a linear decay to zero. Warmup lasted for 6% of the full training duration. Global batch size was set to 4096, and microbatch size was 128; since global batch size was 4096, full pretraining consisted of 70,000 batches. We set the maximum sequence length during pretraining to 128, and we used the standard embedding dimension of 768. For MosaicBERT, we applied 0.1 dropout to the feedforward layers but no dropout to the FlashAttention module, as this was not possible with the OpenAI triton implementation. Full configuration details for pretraining MosaicBERT-Base can be found in the configuration yamls in the mosaicml/examples repo here.

## Evaluation results

When fine-tuned on downstream tasks (following the finetuning details here), the MosaicBERT model achieves the following GLUE results:

Task MNLI-(m/mm) QQP QNLI SST-2 CoLA STS-B MRPC RTE Average
0.8495 0.9029 0.9074 0.9246 0.5511 0.8927 0.9003 0.8136 0.8428

Note that this is averaged over n=5 pretraining seeds.

## Intended uses & limitations

This model is intended to be finetuned on downstream tasks.