Zhilin Wang, Jiaqi Zeng, Olivier Delalleau, Daniel Egert, Ellie Evans, Hoo-Chang Shin, Felipe Soares, Yi Dong, Oleksii Kuchaiev. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Preference datasets are essential for training general-domain, instruction-following language models with Reinforcement Learning from Human Feedback (RLHF). Each subsequent data release raises expectations for future data collection, meaning there is a constant need to advance the quality and diversity of openly available preference data. To address this need, we introduce HelpSteer3-Preference, a permissively licensed (CC-BY-4.0), high-quality, human-annotated preference dataset comprising of over 40,000 samples. These samples span diverse real-world applications of large language models (LLMs), including tasks relating to STEM, coding and multilingual scenarios. Using HelpSteer3-Preference, we train Reward Models (RMs) that achieve top performance on RM-Bench (82.4%) and JudgeBench (73.7%). This represents a substantial improvement (~10% absolute) over the previously best-reported results from existing RMs. We demonstrate HelpSteer3-Preference can also be applied to train Generative RMs and how policy models can be aligned with RLHF using our RMs.

QE Viewer is a web-based tool for visualizing results of a Machine Translation Quality Estimation (QE) system. It allows users to see information on the predicted post-editing distance (PED) for a given file or sentence, and highlighted words that were predicted to contain MT errors. The tool can be used in a variety of academic, educational and commercial scenarios.

The Google Patents is one of the main important sources of patents information. A striking characteristic is that many of its abstracts are presented in more than one language, thus making it a potential source of parallel corpora. This article presents the development of a parallel corpus from the open access Google Patents dataset in 74 language pairs, comprising more than 68 million sentences and 800 million tokens. Sentences were automatically aligned using the Hunalign algorithm for the largest 22 language pairs, while the others were abstract (i.e. paragraph) aligned. We demonstrate the capabilities of our corpus by training Neural Machine Translation (NMT) models for the main 9 language pairs, with a total of 18 models. Our parallel corpus is freely available in TSV format and with a SQLite database, with complementary information regarding patent metadata.