YuLan-Mini: An Open Data-efficient Language Model
This work addresses the problem of high computational costs in LLM development for researchers and practitioners, presenting an incremental improvement in data efficiency.
The paper tackles the challenge of resource-intensive pre-training for large language models by introducing YuLan-Mini, a 2.42B parameter model that achieves top-tier performance among similar-scale models, trained on only 1.08T tokens to match industry leaders using more data.
Effective pre-training of large language models (LLMs) has been challenging due to the immense resource demands and the complexity of the technical processes involved. This paper presents a detailed technical report on YuLan-Mini, a highly capable base model with 2.42B parameters that achieves top-tier performance among models of similar parameter scale. Our pre-training approach focuses on enhancing training efficacy through three key technical contributions: an elaborate data pipeline combines data cleaning with data schedule strategies, a robust optimization method to mitigate training instability, and an effective annealing approach that incorporates targeted data selection and long context training. Remarkably, YuLan-Mini, trained on 1.08T tokens, achieves performance comparable to industry-leading models that require significantly more data. To facilitate reproduction, we release the full details of the data composition for each training phase. Project details can be accessed at the following link: https://github.com/RUC-GSAI/YuLan-Mini.