Ziya2: Data-centric Learning is All LLMs Need

TL;DR

This paper argues that encoder-only models can achieve strong natural language understanding while reducing training costs compared to decoder-heavy LLMs. It presents UniMC-1.3B-Base, an encoder-only model built on an LLama-compatible architecture, trained with a two-stage regime on 600B unsupervised tokens plus millions of supervised examples, and a 4096-token RoPE-enabled context. Through extensive data curation (deduplication and quality scoring) and targeted fine-tuning, the approach yields competitive results on English, Chinese, and code benchmarks, rivaling some decoder-only models. The work highlights the practical value of data-centric strategies in scaling and stabilizing encoder-only models for broad NLP tasks.

Abstract

Various large language models (LLMs) have been proposed in recent years, including closed- and open-source ones, continually setting new records on multiple benchmarks. However, the development of LLMs still faces several issues, such as high cost of training models from scratch, and continual pre-training leading to catastrophic forgetting, etc. Although many such issues are addressed along the line of research on LLMs, an important yet practical limitation is that many studies overly pursue enlarging model sizes without comprehensively analyzing and optimizing the use of pre-training data in their learning process, as well as appropriate organization and leveraging of such data in training LLMs under cost-effective settings. In this work, we propose Ziya2, a model with 13 billion parameters adopting LLaMA2 as the foundation model, and further pre-trained on 700 billion tokens, where we focus on pre-training techniques and use data-centric optimization to enhance the learning process of Ziya2 on different stages. We define three data attributes and firstly establish data-centric scaling laws to illustrate how different data impacts LLMs. Experiments show that Ziya2 significantly outperforms other models in multiple benchmarks especially with promising results compared to representative open-source ones. Ziya2 (Base) is released at https://huggingface.co/IDEA-CCNL/Ziya2-13B-Base and https://modelscope.cn/models/Fengshenbang/Ziya2-13B-Base/summary.