High-quality instruction-tuning data is crucial for developing Large Language Models (LLMs) that can effectively navigate real-world tasks and follow human instructions. While synthetic data generation offers a scalable approach for creating such datasets, it imposes a quality ceiling where models trained on the data cannot outperform the LLM generating it. To overcome this limitation, we introduce Reference-Level Feedback, a paradigm that extracts desirable characteristics from carefully curated reference samples to guide the synthesis of higher-quality instruction-response pairs. Using this approach, we synthesize REFED, a dataset of 10K instruction-response pairs. Fine-tuning Llama-3.1-8B-Instruct and Mistral-7B-Instruct on REFED demonstrate state-of-the-art performance among similarly sized models, notably reaching a 43.96% length-controlled win-rate on AlpacaEval 2.0. Extensive experiments demonstrate that Reference-Level Feedback consistently outperforms traditional sample-level feedback methods, generalizes across model architectures, and produces high-quality and diverse data at low cost.

Large Language Models (LLMs) have demonstrated remarkable performance across various NLP tasks, yet they continue to face challenges in discourse relation recognition (DRR). Current state-of-the-art methods for DRR primarily rely on smaller pre-trained language models (PLMs). In this study, we conduct a comprehensive analysis of different approaches using both PLMs and LLMs, evaluating their effectiveness for DRR at multiple granularities and under different data availability settings. Our findings indicate that no single approach consistently outperforms the others, and we offer a general comparison framework to guide the selection of the most appropriate model based on specific DRR requirements and data conditions.

Automatic evaluation of natural language generation has long been an elusive goal in NLP. A recent paradigm fine-tunes pre-trained language models to emulate human judgements for a particular task and evaluation criterion. Inspired by the generalization ability of instruction-tuned models, we propose a learned metric based on instruction tuning. To test our approach, we collected HEAP, a dataset of human judgements across various NLG tasks and evaluation criteria. Our findings demonstrate that instruction tuning language models on HEAP yields good performance on many evaluation tasks, though some criteria are less trivial to learn than others. Further, jointly training on multiple tasks can yield additional performance improvements, which can be beneficial for future tasks with little to no human annotated data.