It's All in The [MASK]: Simple Instruction-Tuning Enables BERT-like Masked Language Models As Generative Classifiers

TL;DR

This paper shows that encoder-only models like BERT can be used as generative classifiers with simple training and inference, achieving competitive zero-shot and fine-tuned performance without complex prompts or architecture changes.

Contribution

Introducing ModernBERT-Large-Instruct, a 0.4B-parameter encoder model that leverages its MLM head for generative classification, outperforming similar-sized LLMs and matching traditional classifiers.

Findings

Strong zero-shot performance on classification and knowledge tasks.

Generative approach matches or surpasses traditional classifiers after fine-tuning.

Performance depends on training data diversity and volume.

Abstract

While encoder-only models such as BERT and ModernBERT are ubiquitous in real-world NLP applications, their conventional reliance on task-specific classification heads can limit their applicability compared to decoder-based large language models (LLMs). In this work, we introduce ModernBERT-Large-Instruct, a 0.4B-parameter encoder model that leverages its masked language modelling (MLM) head for generative classification. Our approach employs an intentionally simple training loop and inference mechanism that requires no heavy pre-processing, heavily engineered prompting, or architectural modifications. ModernBERT-Large-Instruct exhibits strong zero-shot performance on both classification and knowledge-based tasks, outperforming similarly sized LLMs on MMLU and achieving 93% of Llama3-1B's MMLU performance with 60% less parameters. We also demonstrate that, when fine-tuned, the generative approach using the MLM head matches or even surpasses traditional classification-head methods across diverse NLU tasks.This capability emerges specifically in models trained on contemporary, diverse data mixes, with models trained on lower volume, less-diverse data yielding considerably weaker performance. Although preliminary, these results demonstrate the potential of using the original generative masked language modelling head over traditional task-specific heads for downstream tasks. Our work suggests that further exploration into this area is warranted, highlighting many avenues for future improvements.