Fine-Tuning Causal LLMs for Text Classification: Embedding-Based vs. Instruction-Based Approaches

We explore efficient strategies to fine-tune decoder-only Large Language Models (LLMs) for downstream text classification under resource constraints. Two approaches are investigated: (1) attaching a classification head to a pre-trained causal LLM and fine-tuning on the task (using the LLM's final token embedding as a sequence representation), and (2) instruction-tuning the LLM in a prompt->response format for classification. To enable single-GPU fine-tuning of models up to 8B parameters, we combine 4-bit model quantization with Low-Rank Adaptation (LoRA) for parameter-efficient training. Experiments on two datasets - a proprietary single-label dataset and the public WIPO-Alpha patent dataset (extreme multi-label classification) - show that the embedding-based method significantly outperforms the instruction-tuned method in F1-score, and is very competitive with - even surpassing - fine-tuned domain-specific models (e.g. BERT) on the same tasks. These results demonstrate that directly leveraging the internal representations of causal LLMs, along with efficient fine-tuning techniques, yields impressive classification performance under limited computational resources. We discuss the advantages of each approach while outlining practical guidelines and future directions for optimizing LLM fine-tuning in classification scenarios.

Patent Language Model Pretraining with ModernBERT

Transformer-based language models such as BERT have become foundational in NLP, yet their performance degrades in specialized domains like patents, which contain long, technical, and legally structured text. Prior approaches to patent NLP have primarily relied on fine-tuning general-purpose models or domain-adapted variants pretrained with limited data. In this work, we pretrain 3 domain-specific masked language models for patents, using the ModernBERT architecture and a curated corpus of over 60 million patent records. Our approach incorporates architectural optimizations, including FlashAttention, rotary embeddings, and GLU feed-forward layers. We evaluate our models on four downstream patent classification tasks. Our model, ModernBERT-base-PT, consistently outperforms the general-purpose ModernBERT baseline on three out of four datasets and achieves competitive performance with a baseline PatentBERT. Additional experiments with ModernBERT-base-VX and Mosaic-BERT-large demonstrate that scaling the model size and customizing the tokenizer further enhance performance on selected tasks. Notably, all ModernBERT variants retain substantially faster inference over - 3x that of PatentBERT - underscoring their suitability for time-sensitive applications. These results underscore the benefits of domain-specific pretraining and architectural improvements for patent-focused NLP tasks.

End-to-End Document Classification and Key Information Extraction using Assignment Optimization

We propose end-to-end document classification and key information extraction (KIE) for automating document processing in forms. Through accurate document classification we harness known information from templates to enhance KIE from forms. We use text and layout encoding with a cosine similarity measure to classify visually-similar documents. We then demonstrate a novel application of mixed integer programming by using assignment optimization to extract key information from documents. Our approach is validated on an in-house dataset of noisy scanned forms. The best performing document classification approach achieved 0.97 f1 score. A mean f1 score of 0.94 for the KIE task suggests there is significant potential in applying optimization techniques. Abation results show that the method relies on document preprocessing techniques to mitigate Type II errors and achieve optimal performance.

Unimodal and Multimodal Representation Training for Relation Extraction

Multimodal integration of text, layout and visual information has achieved SOTA results in visually rich document understanding (VrDU) tasks, including relation extraction (RE). However, despite its importance, evaluation of the relative predictive capacity of these modalities is less prevalent. Here, we demonstrate the value of shared representations for RE tasks by conducting experiments in which each data type is iteratively excluded during training. In addition, text and layout data are evaluated in isolation. While a bimodal text and layout approach performs best (F1=0.684), we show that text is the most important single predictor of entity relations. Additionally, layout geometry is highly predictive and may even be a feasible unimodal approach. Despite being less effective, we highlight circumstances where visual information can bolster performance. In total, our results demonstrate the efficacy of training joint representations for RE.