Pfeiffer, J., RΓΌcklΓ©, A., Poth, C., Kamath, A., Vulic, I., Ruder, S., Cho, K., & Gurevych, I. (2020). AdapterHub: A Framework for Adapting Transformers. ArXiv, abs/2007.07779.

Transformer models pre-trained on massive amounts of text data and subsequently fine-tuned on target tasks have led to considerable advances in NLP, achieving state-of-the-art results across the board. However, models such as BERT (Devlin et al., 2019) and RoBERTa (Liu et al., 2019) consist of several millions of parameters, and thus, sharing and distributing fully fine-tuned models for each individual downstream task can be prohibitive.

Adapters are a light-weight alternative to full model fine-tuning, consisting of only a tiny set of newly introduced parameters at every transformer layer. Adapters overcome several limitations typically observed with full model fine-tuning: they are parameter-efficient, they speed up training iterations, and they are shareable and composable due to their modularity and compact size. Moreover, adapters usually perform on-par with state-of-the-art full fine-tuning.

With multiple different adapter architectures and a wide variety of pre-trained transformers available, training, sharing and re-using adapters is not straightforward. As a solution, AdapterHub, A Framework for Adapting Transformers provides a unified interface to different adapter architectures and composition techniques, making them widely accessible to the research community. Built on top of Huggingface's Transformers framework, the AdapterHub has access to a large base of pre-trained transformers. In the following, we will go through the process of training, sharing, and composing adapters with AdapterHub.

A Short Introduction to Adapters

Figure 1: Steps of working with adapters

Adapters provide a lightweight alternative to fully fine-tuning a pre-trained language model on a downstream task. For a transformer-based architecture, a small set of new parameters is introduced in every transformer layer. While different adapter architectures are possible, a simple layout using a down- and an up-projection layer first introduced by Houlsby et al. (2020) has proven to work well (see Figure 1 for illustration). In many cases, adapters perform on-par with fully fine-tuned models.

During training on the target task, all weights of the pre-trained language model are kept fix. The only weights to be updated are those introduced by the adapter modules. This results in modular knowledge representations which subsequently can be easily extracted from the underlying language model. The extracted adapter modules then can be distributed independently and plugged into a language model dynamically. The encapsulated character of adapters also allows for easy exchange and composition of different adapters (Pfeiffer et al., 2020a). Since this workflow of using adapters is very universal, it can potentially be applied to a wide range of different use cases. As an example, adapters have been used successfully for zero-shot cross-lingual transfer between different tasks (Pfeiffer et al., 2020b). Figure 1 illustrates the described adapter workflow.

Figure 2: Size comparison of a fully fine-tuned model and an adapter

Using adapters provides various benefits, especially in parameter efficiency. The amount of updated adapter parameters are only about 1% of the fully fine-tuned model and in many cases only requires a few Megabytes of storage space. This makes it easy to share adapters, store adapters for many different tasks and load additional adapters on-the-fly. Additionally, their compact size with the majority of weights bein frozen, makes adapters a computationally efficient fine-tuning choice (RΓΌcklΓ© et al., 2020).

What is AdapterHub?

With AdapterHub, we have developed a framework which makes working with adapters straightforward. AdapterHub is divided into two core components: adapter-transformers, a library built on top of HuggingFace transformers that integrates adapter support into various popular Transformer-based language models, and the Hub, an open platform for sharing, exploring and consuming pre-trained adapters.

Figure 3: The AdapterHub lifecycle

Based on Figure 3, we'll go through the lifecycle of working with AdapterHub on a higher level: HuggingFace transformers (πŸ€—) builds the backbone of our framework. A user who wants to train an adapter (πŸ‘©πŸΎβ€πŸ’») loads a pre-trained language model (πŸ€–) from πŸ€—. In β‘ , new adapter modules are introduced to the loaded language model. Afterwards, πŸ‘©πŸΎβ€πŸ’» trains the adapter on a downstream task (β‘‘). As soon as training has completed, πŸ‘©πŸΎβ€πŸ’» can extract the trained adapter weights from the (unaltered) πŸ€– in β‘’. πŸ‘©πŸΎβ€πŸ’» packs the adapter weights and uploads them to the Hub. Here, πŸ‘¨πŸΌβ€πŸ’» can find the pre-trained adapter in step β‘£. Together with downloading the matching πŸ€– from πŸ€—, πŸ‘¨πŸΌβ€πŸ’» then can download the adapter from the Hub and integrate it into his own model (β‘€). In β‘₯, he lastly can apply πŸ‘©πŸΎβ€πŸ’»'s adapter for his own purposes.

In the following, we will have a look at some of these steps in more detail.

Training an Adapter

Open In Colab

Training an adapter on a downstream task is a straightforward process using adapter-transformers which can be installed via pip:

pip install adapter-transformers

This package is fully compatible with HuggingFace's transformers library and can act as a drop-in replacement. Therefore, we can instantiate a pre-trained language model and tokenizer in the familiar way:

from transformers import RobertaTokenizer, RobertaConfig, RobertaModelWithHeads

tokenizer = RobertaTokenizer.from_pretrained(
config = RobertaConfig.from_pretrained(
    id2label={ 0: "πŸ‘Ž", 1: "πŸ‘"},
model = RobertaModelWithHeads.from_pretrained(

There is one difference compared to HuggingFace transformers in the code above: We use the new class RobertaModelWithHeads which allows a more flexible way of configuring prediction heads.

The next steps configure our adapter setup. Note that these are the only lines additionally needed to switch from full fine-tuning to adapter training.

from transformers import AdapterType

# Add a new adapter
model.add_adapter("rotten_tomatoes", AdapterType.text_task)
# Add a matching classification head
model.add_classification_head("rotten_tomatoes", num_labels=2)
# Activate the adapter

We add a new adapter to our model by calling add_adapter(). We pass a name ("rotten_tomatoes") and the type of adapter (task adapter). Next, we add a binary classification head. It's convenient to give the prediction head the same name as the adapter. This allows us to activate both together in the next step. The train_adapter() method does two things:

  1. It freezes all weights of the pre-trained model so only the adapter weights are updated during training.
  2. It activates the adapter and the prediction head such that both are used in every forward pass.

All the rest of the training process is identical to a full fine-tuning approach. Check out the Colab notebook on adapter training to see the full code.

In the end, the trained adapter can be exported to the file system using a single line of code:

model.save_adapter("./final_adapter", "rotten_tomatoes")

Interacting with the Hub

Open In Colab

The adapter weights trained in the previous section subsequently can be distributed via the Hub, the second core component of the AdapterHub framework. The Hub infrastructure is based on plain YAML description files contributed to a central GitHub repository. The full process of contributing pre-trained adapters is described in our documentation.

The Explore section of the AdapterHub website acts as the starting point for discovering and consuming available pre-trained adapters. A matching adapter can be selected by task domain, training dataset, model architecture and adapter architecture and loaded into adapter-transformers in the following.

Before loading the adapter, we instantiate the model we want to use, a pre-trained bert-base-uncased model from HuggingFace.

from transformers import AutoTokenizer, AutoModelWithHeads

tokenizer = AutoTokenizer.from_pretrained("bert-base-uncased")
model = AutoModelWithHeads.from_pretrained("bert-base-uncased")

Using load_adapter(), we download and add a pre-trained adapter from the Hub. The first parameter specifies the name of the adapter whereas the second selects the adapter architectures to search for.

Also note that most adapters come with a prediction head included. Thus, this method will also load the question answering head trained together with the adapter.

adapter_name = model.load_adapter("qa/squad1@ukp", config="houlsby")

With set_active_adapters() we tell our model to use the adapter we just loaded in every forward pass.


Again, these are all changes needed to set up a pre-trained language model with a pre-trained adapter. The rest of the inference is identical to a setup without adapters. To see a full example, check out the Colab notebook for adapter inference.

Adapter Composition

Open In Colab

As presented earlier, adapters are especially suitable for various kinds of compositions on a new target task. One of these composition approaches is AdapterFusion (Pfeiffer et al., 2020) which is also tightly integrated into AdapterHub.

The knowledge learned by multiple pre-trained adapters from the Hub can be leveraged to solve a new target task. In this setup, only a newly introduced fusion layer is trained while the rest of the model is kept fix.

First, we load three adapters pre-trained on different tasks from the Hub: MultiNLI, QQP and QNLI. As we don't need their prediction heads, we pass with_head=False to the loading method. Next, we add a new fusion layer that combines all the adapters we've just loaded. Finally, we add a new classification head for our target task on top.

from transformers import AdapterType

# Load the pre-trained adapters we want to fuse
model.load_adapter("nli/multinli@ukp", AdapterType.text_task, load_as="multinli", with_head=False)
model.load_adapter("sts/qqp@ukp", AdapterType.text_task, with_head=False)
model.load_adapter("nli/qnli@ukp", AdapterType.text_task, with_head=False)
# Add a fusion layer for all loaded adapters
model.add_fusion(["multinli", "qqp", "qnli"])

# Add a classification head for our target task
model.add_classification_head("cb", num_labels=len(id2label))

The last preparation step is to define and activate our adapter setup. Similar to train_adapter(), train_fusion() does two things: It freezes all weights of the model (including adapters!) except for the fusion layer and classification head. It also activates the given adapter setup to be used in very forward pass.

The syntax for the adapter setup (which is also applied to other methods such as set_active_adapters()) works as follows:

  • a single string is interpreted as a single adapter
  • a list of strings is interpreted as a stack of adapters
  • a nested list of strings is interpreted as a fusion of adapters
# Unfreeze and activate fusion setup
adapter_setup = [
  ["multinli", "qqp", "qnli"]

See the full training example in the Colab notebook on AdapterFusion.


Adapters are a promising new approach to transfer learning in NLP, providing benefits in efficiency and modularity. AdapterHub provides tools for the full lifecycle of interacting with adapters. The integration into the successful HuggingFace transformers framework makes it straightforward to adapt training setups to adapters. AdapterHub is continuously evolving with the addition of adapter support to new models, the integration of new application scenarios for adapters and a growing platform of pre-trained adapter modules.


  • Devlin, J., Chang, M., Lee, K., & Toutanova, K. (2019). BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. NAACL-HLT.
  • Houlsby, N., Giurgiu, A., Jastrzebski, S., Morrone, B., Laroussilhe, Q.D., Gesmundo, A., Attariyan, M., & Gelly, S. (2019). Parameter-Efficient Transfer Learning for NLP. ICML.
  • Liu, Y., Ott, M., Goyal, N., Du, J., Joshi, M., Chen, D., Levy, O., Lewis, M., Zettlemoyer, L., & Stoyanov, V. (2019). RoBERTa: A Robustly Optimized BERT Pretraining Approach. ArXiv, abs/1907.11692.
  • Pfeiffer, J., Kamath, A., RΓΌcklΓ©, A., Cho, K., & Gurevych, I. (2020). AdapterFusion: Non-Destructive Task Composition for Transfer Learning. ArXiv, abs/2005.00247.
  • Pfeiffer, J., Vulic, I., Gurevych, I., & Ruder, S. (2020). MAD-X: An Adapter-based Framework for Multi-task Cross-lingual Transfer. ArXiv, abs/2005.00052.
  • RΓΌcklΓ©, A., Geigle, G., Glockner, M., Beck, T., Pfeiffer, J., Reimers, N., & Gurevych, I. (2020). AdapterDrop: On the Efficiency of Adapters in Transformers. ArXiv, abs/2010.11918.
  • Wolf, T., Debut, L., Sanh, V., Chaumond, J., Delangue, C., Moi, A., Cistac, P., Rault, T., Louf, R., Funtowicz, M., & Brew, J. (2019). HuggingFace's Transformers: State-of-the-art Natural Language Processing. ArXiv, abs/1910.03771.