Yingxin Lin | Searcharxiv

arXiv2026

The First ChineseBabyLM Challenge: training data-efficient and cognitively plausible language models for Chinese

This paper presents the first ChineseBabyLM Challenge, organized as part of NLPCC 2026. The challenge asked participants to train language models from scratch using no more than 102M Chinese words. The models were evaluated on three tracks: natural language understanding, cognitive alignment, and Hanzi knowledge. There were no restrictions on tokenizers, model architectures, or the number of training epochs. Eighteen teams submitted 28 distinct models, generating 74 result files. The overall-winning team used a DeBERTa-v2 architecture and introduced an auxiliary pinyin-prediction objective during pretraining. Several submissions also explored curriculum-learning strategies and architectural innovations. Overall, the challenge provides a benchmark for advancing data-efficient and cognitively plausible approaches to Chinese language modeling.

cs.CL↗

arXiv2024

False Discovery Rate Control via Data Splitting for Testing-after-Clustering

Testing for differences in features between clusters in various applications often leads to inflated false positives when practitioners use the same dataset to identify clusters and then test features, an issue commonly known as ``double dipping''. To address this challenge, inspired by data-splitting strategies for controlling the false discovery rate (FDR) in regressions \parencite{daiFalseDiscoveryRate2023}, we present a novel method that applies data-splitting to control FDR while maintaining high power in unsupervised clustering. We first divide the dataset into two halves, then apply the conventional testing-after-clustering procedure to each half separately and combine the resulting test statistics to form a new statistic for each feature. The new statistic can help control the FDR due to its property of having a sampling distribution that is symmetric around zero for any null feature. To further enhance stability and power, we suggest multiple data splitting, which involves repeatedly splitting the data and combining results. Our proposed data-splitting methods are mathematically proven to asymptotically control FDR in Gaussian settings. Through extensive simulations and analyses of single-cell RNA sequencing (scRNA-seq) datasets, we demonstrate that the data-splitting methods are easy to implement, adaptable to existing single-cell data analysis pipelines, and often outperform other approaches when dealing with weak signals and high correlations among features.

stat.ME↗