Finally, a Replacement for BERT

BERT was released in 2018 (millennia ago in AI-years!) and yet it’s still widely used today: in fact, it’s currently the second most monthly downloaded model on the HuggingFace hub, with more than 68 million monthly downloads, only second to another encoder model fine-tuned for retrieval. That’s because its encoder-only architecture makes it ideal for the kinds of real-world problems that come up every day, like retrieval (such as for RAG), classification (such as content moderation), and entity extraction (such as for privacy and regulatory compliance).

Finally, 6 years later, we have a replacement! Today, we at Answer.AI and LightOn (and friends!) are releasing ModernBERT. ModernBERT is a new model series that is a Pareto improvement over BERT and its younger siblings across both speed and accuracy. This model takes dozens of advances from recent years of work on large language models (LLMs), and applies them to a BERT-style model, including updates to the architecture and the training process.

We expect to see ModernBERT become the new standard in the numerous applications where encoder-only models are now deployed, such as in RAG pipelines (Retrieval Augmented Generation) and recommendation systems.

In addition to being faster and more accurate, ModernBERT also increases context length to 8k tokens (compared to just 512 for most models), and is the first encoder-only model that includes a large amount of code in its training data. These features open up new application areas that were previously inaccessible through open models, such as large-scale code search, new IDE features, and new types of retrieval pipelines based on full document retrieval rather than small chunks.

But in order to explain just what we did, let’s first take a step back and look at where we’ve come from.

Decoder-only models

The recent high-profile advances in LLMs have been in models like GPT, Llama, and Claude. These are decoder-only models, or generative models. Their ability to generate human-like content has enabled astonishing new GenAI application areas like generated art and interactive chat. These striking applications have attracted major investment, funded booming research, and led to rapid technical advances. What we’ve done, essentially, is port these advances back to an encoder-only model.

Why? Because many practical applications need a model that’s lean and mean! And it doesn’t need to be a generative model.

More bluntly, decoder-only models are too big, slow, private, and expensive for many jobs. Consider that the original GPT-1 was a 117 million parameter model. The Llama 3.1 model, by contrast, has 405 billion parameters, and its technical report describes a data synthesis and curation recipe that is too complex and expensive even for most corporations to reproduce. So to use such a model, like ChatGPT, you pay in cents and wait in seconds to get an API reply back from heavyweight servers outside of your control.

Of course, the open-ended capabilities of these giant generative models mean that you can, in a pinch, press them into service for non-generative or discriminative tasks, such as classification. This is because you can describe a classification task in plain English and ... just ask the model to classify. But while this workflow is great for prototyping, you don’t want to pay prototype prices once you’re in mass production.

The popular buzz around GenAI has obscured the role of encoder-only models. These are the workhorses of practical language processing, the models that are actually being used for such workloads right now in many scientific and commercial applications.

‍

Encoder-only models

The output of an encoder-only model is a list of numerical values (an embedding vector). You might say that instead of answering with text, an encoder model literally encodes its “answer” into this compressed, numerical form. That vector is a compressed representation of the model's input, which is why encoder-only are sometimes referred to as representational models. 

While decoder-only models (like a GPT) can do the work of an encoder-only model (like a BERT), they are hamstrung by a key constraint: since they are generative models, they are mathematically “not allowed” to “peek” at later tokens. They can only ever look backwards. This is in contrast to encoder-only models, which are trained so each token can look forwards and backwards (bi-directional). They are built for this, and it makes them very efficient at what they do.

Basically, a frontier model like OpenAI's O1 is like a Ferrari SF-23. It’s an obvious triumph of engineering, designed to win races, and that’s why we talk about it. But it takes a special pit crew just to change the tires and you can’t buy one for yourself. In contrast, a BERT model is like a Honda Civic. It’s also an engineering triumph, but more subtly, since it is engineered to be affordable, fuel-efficient, reliable, and extremely useful. And that’s why they’re absolutely everywhere.

You can see this by looking at it a number of ways.

Supporting generative models: One way to understand the prevalence of representational models (encoder-only) is to note how frequently they are used in concert with a decoder-only model to make a system which is safe and efficient.

The obvious example is RAG. Instead of relying on the LLM’s knowledge trained into the model’s parameters, the system uses a document store to furnish the LLM with information relevant to the query. But of course this only defers the problem. If the LLM doesn’t know which documents are relevant to the query, then the system will need some other process to select those documents? It’s going to need a model which is fast and cheap enough that it can be used to encode the large quantities of information needed to make the LLM useful. That model is often a BERT-like encoder-only model.

Another example is supervision architectures, where a cheap classifier might be used to ensure that generated text does not violate content safety requirements.

In short, whenever you see a decoder-only model in deployment, there’s a reasonable chance an encoder-only model is also part of the system. But the converse is not true.

Encoder-based systems: Before there was GPT, there were content recommendations in social media and in platforms like Netflix. There was ad targeting in those venues, in search, and elsewhere. There was content classification for spam detection, abuse detection, etc.. These systems were not built on generative models, but on representational models like encoder-only models. And all these systems are still out there and still running at enormous scale. Imagine how many ads are targeted per second around the world!

Downloads: On HuggingFace, RoBERTa, one of the leading BERT-based models, has more downloads than the 10 most popular LLMs on HuggingFace combined. In fact, currently, encoder-only models add up to over a billion downloads per month, nearly three times more than decoder-only models with their 397 million monthly downloads. In fact, the `fill-mask` model category, composed of encoder “base models” such as ModernBERT, ready to be fine-tuned for other downstream applications, is the most downloaded model category overall.

Inference costs: What the above suggests, is that on an inference-per-inference basis, there are many times more inferences performed per year on encoder-only models than on decoder-only or generative models. An interesting example is FineWeb-Edu, where model-based quality filtering had to be performed over 15 trillion tokens. The FineWeb-Edu team chose to generate annotations with a decoder-only model, Llama-3-70b-Instruct, and perform the bulk of the filtering with a fine-tuned BERT-based model. This filtering took 6,000 H100 hours, which, at HuggingFace Inference Points’ pricing of $10/hour, comes to a total of $60,000. On the other hand, feeding 15 trillion tokens to popular decoder-only models, even with the lowest-cost option of using Google’s Gemini Flash and its low inference cost of $0.075/million tokens, would cost over one million dollars!

‍

Performance

Overview

Here’s a snapshot of the accuracy of ModernBERT and other models across a range of tasks, as measured by standard academic benchmarks – as you can see, ModernBERT is the only model which is a top scorer across every category, which makes it the one model you can use for all your encoder-based tasks:

If you’ve ever done an NLP competition on Kaggle, then you’ll know that DeBERTaV3 has been the choice of champions for years. But no longer: not only is ModernBERT the first base-size model to beat DeBERTaV3 on GLUE, it also uses less than 1/5th of Deberta’s memory.

And of course, ModernBERT is fast. It’s twice as fast as DeBERTa – in fact, up to 4x faster in the more common situation where inputs are mixed length. Its long context inference is nearly 3 times faster than other high-quality models such as NomicBERT and GTE-en-MLM. 

ModernBERT’s context length of 8,192 tokens is over 16x larger than most existing encoders. This is critical, for instance, in RAG pipelines, where a small context often makes chunks too small for semantic understanding. ModernBERT is also the state-of-the-art long context retriever with ColBERT, and is 9 percentage points above the other long context models. Even more impressive: this very quickly trained model, simply tuned to compare to other backbones, outperforms even widely-used retrieval models on long-context tasks!

For code retrieval, ModernBERT is unique. There’s nothing to really compare it to, since there’s never been an encoder model like this trained on a large amount of code data before. For instance, on the StackOverflow-QA dataset (SQA), which is a hybrid dataset mixing both code and natural language, ModernBERT's specialized code understanding and long-context capabilities make it the only backbone to score over 80 on this task.

This means whole new applications are likely to be built on this capability. For instance, imagine an AI-connected IDE which had an entire enterprise codebase indexed with ModernBERT embeddings, providing fast long context retrieval of the relevant code across all repositories. Or a code chat service which described how an application feature worked that integrated dozens of separate projects.

Compared to the mainstream models, ModernBERT performs better across nearly all three broad task categories of retrieval, natural language understanding, and code retrieval. Whilst it slightly lags DeBERTaV3 in one area (natural language understanding), it is many times faster. Please note that ModernBERT, as any other base model, can only do masked word prediction out-of-the-box. To be able to perform other tasks, the base model should be fine-tuned as done in these boilerplates.

Compared to the specialized models, ModernBERT is comparable or superior in most tasks. In addition, ModernBERT is faster than most models across most tasks, and can handle inputs up to 8,192 tokens, 16x longer than the mainstream models. 

‍

Efficiency

Here’s the memory (max batch size, BS) and Inference (in thousands of tokens per second) efficiency results on an NVIDIA RTX 4090 for ModernBERT and other decoder models:

The first thing you might notice is that we’re analysing the efficiency on an affordable consumer GPU, rather than the latest unobtainable hyped hardware. ModernBERT is focused on, first and foremost, practicality, not hype.

As part of this focus, it also means we’ve made sure it works well for real-world applications, rather than just benchmarks. Models of this kind are normally tested on just the one exact size they’re best at – their maximum context length. That’s what the “fixed” column in the table shows. But in the real world input sizes vary, so that’s the performance we worked hard to optimise – the “variable” column. As you can see, for variable length inputs, ModernBERT is much faster than all other models.

For long context inputs, which we believe will be the basis for the most valuable and important future applications, ModernBERT is 2-3x faster than the next fastest model. And, on the “practicality” dimension again: ModernBERT doesn’t require the additional heavy “xformers” dependency, but instead only requires the now commonplace Flash Attention as a dependency.

Furthermore, thanks to ModernBERT’s efficiency, it can use a larger batch size than nearly any other model, and can be used effectively on smaller and cheaper GPUs. The efficiency of the base size, in particular, may enable new applications that run directly in browsers, on phones, and so forth.

‍

Why is ModernBERT, well, Modern?

Now, we’ve made our case to why we should give some more love to encoder models. As trusted, under-appreciated workhorses, they’ve had surprisingly few updates since BERT, which was released in 2018 - ancient history in AI terms!

Even more surprising: since RoBERTa, there has been no encoder providing overall improvements without tradeoffs (fancily known as “Pareto improvements”): DeBERTaV3 had better GLUE and classification performance, but sacrificed both efficiency and retrieval. Other models, such as AlBERT, or newer ones, like GTE-en-MLM, all improved over the original BERT and RoBERTa in some ways but regressed in others.

However, since the duo’s original release, we've learned an enormous amount about how to build better language models. If you’ve used LLMs at all, you’re very well aware of it: while they’re rare in the encoder-world, Pareto improvements are constant in decoder-land, where models constantly become better at everything. And as we’ve all learned by now: model improvements are only partially magic, and mostly engineering.

The goal of the (hopefully aptly named) ModernBERT project was thus fairly simple: bring this modern engineering to encoder models. We did so in three core ways:

1. a modernized transformer architecture
2. particular attention to efficiency
3. modern data scales & sources

Meet the New Transformer, Same as the Old Transformer

The Transformer architecture has become dominant, and is used by the vast majority of models nowadays. However, it’s important to remember that there isn’t one but many Transformers. The main thing they share in common is their deep belief that attention is indeed all you need, and as such, build various improvements centered around the attention mechanism.

ModernBERT takes huge inspiration from the Transformer++ (as coined by Mamba), first used by the Llama2 family of models. Namely, we replace older BERT-like building blocks with their improved equivalent, namely, we:

• Replace the old positional encoding with "rotary positional embeddings" (RoPE): this makes the model much better at understanding where words are in relation to each other, and allows us to scale to longer sequence lengths.
• Switch out the old MLP layers for GeGLU layers, improving on the original BERT’s GeLU activation function.
• Streamline the architecture by removing unnecessary bias terms, letting us spend our parameter budget more effectively
• Add an extra normalization layer after embeddings, which helps stabilize training

Upgrading a Honda Civic for the Race Track

We’ve covered this already: encoders are no Ferraris, and ModernBERT is no exception. However, that doesn’t mean it can’t be fast. When you get on the highway, you generally don’t go and trade in your car for a race car, but rather hope that your everyday reliable ride can comfortably hit the speed limit.

In fact, for all the application cases we mentioned above, speed is essential. Encoders are very popular in uses where they either have to process tons of data, allowing even tiny speed increments to add up very quickly, or where latency is very important, as is the case on RAG. In a lot of situations, encoders are even run on CPU, where efficiency is even more important if we want results in a reasonable amount of time.

As with most things in research, we build while standing on the shoulders of giants, and heavily leverage Flash Attention 2’s speed improvements. Our efficiency improvements rely on three key components: Alternating Attention, to improve processing efficiency, Unpadding and Sequence Packing, to reduce computational waste, and Hardware-Aware Model Design, to maximise hardware utilization.

Global and Local Attention

One of ModernBERT’s most impactful features is Alternating Attention, rather than full global attention. In technical terms, this means that our attention mechanism only attends to the full input every 3 layers (global attention), while all other layers use a sliding window where every token only attends to the 128 tokens nearest to itself (local attention). 

As attention’s computational complexity balloons up with every additional token, this means ModernBERT can process long input sequences considerably faster than any other model.

In practice, it looks like this:

Conceptually, the reason this works is pretty simple: Picture yourself reading a book. For every sentence you read, do you need to be fully aware of the entire plot to understand most of it (full global attention)? Or is awareness of the current chapter enough (local attention), as long as you occasionally think back on its significance to the main plot (global attention)? In the vast majority of cases, it’s the latter.

Unpadding and Sequence Packing

Another core mechanism contributing to ModernBERT’s efficiency is its use for Unpadding and Sequence packing.

In order to be able to process multiple sequences within the same batch, encoder models require them to be the same length, so they can perform parallel computation. Traditionally, we’ve relied on padding to achieve this: figure out which sentence is the longest, and add meaningless tokens (padding tokens) to fill up every other sequence.

While padding solves the problem, it doesn’t do so elegantly: a lot of compute ends up being spent and wasted on padding tokens, which do not contribute any semantic information.

Comparing padding with sequence packing. Sequence packing (‘unpadding’) avoids wasting compute on padding tokens and has more consistent non-padding token counts per batch. Samples are still processed individually through careful masking.

Unpadding solves this issue: rather than keeping these padding tokens, we remove them all, and concatenate them into mini-batches with a batch size of one, avoiding all unnecessary computations. If you’re using Flash Attention, our implementation of unpadding is even faster than previous methods, which heavily relied on unpadding and repadding sequences as they went through the model: we go one step further by introducing our own implementation of unpadding, relying heavily on recent developments in Flash Attention’s RoPE support. This allows ModernBERT to only have to unpad once, and optionally repad sequences after processing, resulting in a 10-20% speedup over previous methods.

To speed up pre-training even further, unpadding is in good company within our model, as we use it in conjunction with sequence packing. Sequence packing here is a logical next step: as we’re concatenating inputs into a single sequence, and GPUs are very good at parallelisation, we want to maximise the computational efficiency we can squeeze out of a single forward model pass. To do so, we use a greedy algorithm to group individual sequences into concatenated ones that are as close to the model’s maximum input length as possible.

Paying Attention to Hardware

Finally, the third facet of ModernBERT’s efficiency is hardware design.

We attempted to balance two insights that have been highlighted by previous research:

1. Deep & Narrow vs Wide & Shallow: Research shows that deeper models with narrower layers, often perform better than shallow models with fewer, wider layers. However, this is a double-edged sword: the deeper the model, the less parallelizable it becomes, and thus, the slower it runs at identical parameter counts.
2. Hardware Efficiency: Model dimensions need to align well with GPU hardware for maximum performance, and different target GPUs result in different constraints.

Sadly, there is no magic recipe to make a model run similarly well on a wide range of GPUs, but there is an excellent cookbook: The Case for Co-Designing Model Architectures with Hardware, in which the ways to optimize a model architecture for a given GPU are carefully laid out. We came up with a heuristic to extend their method to a basket of GPUs, while respecting a given set of constraints. Logically, the first step is to define said constraints, in our case:

• Defining our target GPUs as common inference ones (RTX 3090/4090, A10, T4, L4)
• Roughly defining our target model sizes at 130-to-150 million parameters for ModernBERT-Base, and 350-to-420 for ModernBERT-Large.
• The final embedding sizes must match the original BERT’s dimensions, 768 for base and 1024 for large, to maximize backwards compatibility
• Set performance constraints which are common across the basket of GPUs

Afterwards, we experimented with multiple model designs via a constrained grid search, varying both layer counts and layer width. Once we’d identified shapes that appeared to be the most efficient ones, we confirmed that our heuristics matched real-world GPU performance, and settled on the final model designs.

‍

Training

Data

def data(): return [‘text’, ‘bad_text’, ‘math’, ‘code’]

Picture this exact scene, but replace Developers with Data

Another big aspect in which encoders have been trailing behind is training data. This is often understood to mean solely training data scale, but this is not actually the case: previous encoders, such as DeBERTaV3, were trained for long enough that they might have even breached the trillion tokens scale!

The issue, rather, has been training data diversity: many of the older models train on limited corpora, generally consisting of Wikipedia and Wikibooks. These data mixtures are very noticeably single text modality: they contain nothing but high-quality natural text.
In contrast, ModernBERT is trained on data from a variety of English data, including web documents, code, and scientific articles. It is trained on 2 trillion tokens, of which most are unique, rather than the standard 20-to-40 repetitions common in previous encoders.

The impact of this is immediately noticeable: out of all the existing open source encoders, ModernBERT is in a class of its own on programming-related tasks. We’re particularly interested in what downstream uses this will lead to, in terms of improving programming assistants.

Process

We stick to the original BERT’s training recipe, with some slight upgrades inspired by subsequent work: we remove the Next-Sentence Prediction objective, since then shown to add overhead for no clear gains, and increase the masking rate from 15% to 30%.

Both models are trained with a three-phase process. First, we train on 1.7T tokens at a sequence length of 1024. We then adopt a long-context adaptation phase, training on 250B tokens at a sequence length of 8192, while keeping the total tokens seen per batch more or less consistent by lowering the batch size. Finally, we perform annealing on 50 billion tokens sampled differently, following the long-context extension ideal mix highlighted by ProLong.

Training in three phases is our way of ensuring our model is good across the board, which is reflected in its results: it is competitive on long-context tasks, at no cost to its ability to process short context…

… But it has another benefit: for the first two-phases, we train using a constant learning rate once the warmup phase is complete, and only perform learning rate decay on the final 50 billion tokens, following the Trapezoidal (or Warmup-Stable-Decay) learning rate. And what’s more: we will release every single immediate intermediate checkpoints from these stable phases, inspired by Pythia. Our main reason for doing so was supporting future research and applications: anyone is free to restart training from any of our pre-decay checkpoints, and perform annealing on domain-appropriate data for their intended use!

The tricks, it’s all about the tricks!

If you’ve made it this far into this announcement, you’re probably used to this: of course, we use tricks to make things quicker here too. To be precise, we have two main tricks.

Let’s start with the first one, which is pretty common: since the initial training steps are updating random weights, we adopt batch-size warmup: we start with a smaller batch size so the same number of tokens update the model weights more often, then gradually increase the batch size to the final training size. This significantly speeds up the initial phase of model training, where the model learns its most basic understanding of language.

The second trick is far more uncommon: weight initialization via tiling for the larger model size, inspired by Microsoft’s Phi family of models. This one’s based on the following realization: Why initialize the ModernBERT-large’s initial weights with random numbers when we have a perfectly good (if we dare say so ourselves) set of ModernBERT-base weights just sitting there?

And indeed, it turns out that tiling ModernBERT-base’s weights across ModernBERT-large works better than initializing from random weights. It also has the added benefit of stacking nicely with batch size warmup for even faster initial training.

‍

Conclusion

In this blog post we introduced the ModernBERT models, a new state-of-the-art family of small and efficient encoder-only models, finally giving BERT a much needed do-over.

ModernBERT demonstrates that encoder-only models can be improved by modern methods. They continue to offer very strong performance on some tasks, providing an extremely attractive size/performance ratio.

More than anything, we’re really looking forward to seeing what creative ways to use these models the community will come up with! To encourage this, we’re opening a call for demos until January 10th, 2025: the 5 best ones will get added to this post in a showcase section and win a $100 (or local currency equivalent) Amazon gift card, as well as a 6-month HuggingFace Pro subscription! (If you need a hint to get started, here’s a demo we thought about: code similarity HF space)

‍

Links

• 🤗ModernBERT-Base
• 🤗ModernBERT-Large
• arXiv paper
• 🤗ModernBERT documentation page

The compute for this project was sponsored by LightOn on Orange Business Cloud Avenue.