Resources
Academic
ZhoBLiMP: a systematic assessment of language models with linguistic minimal pairs in Chinese. ZhoBLiMP is a dataset that can be used to probe Mandarin Chinese linguistic knowledge in language models, especially syntax. To validate ZhoBLiMP, we trained 5 Zh-Pythia models from scratch on original Mandarin Chinese data.
- Dataset: ZhoBLiMP covers 118 paradigms in 15 high-level linguistic phenomena. It contains 35k minimal pairs that differ in a minimal way to demonstrate a single syntactic or semantic contrast.
- Models: The parameters of the Zh-Pythia models correspond to their English counterparts (size: 14M, 70M, 160M, 410M, 1.4B).
captain-nemo: a toolkit for running large batches of GPU jobs on the Nautilus (NRP) Kubernetes cluster without tripping its fair-use gate. It paces launches by NRP’s utilization-violation budget, keeps jobs off faulty nodes, checks job manifests for setup steps that can hang, and produces an hourly health report. It ships as a skill that Claude Code and other coding agents can follow; the scripts also run on their own.
- Github: code, quick start, and the lessons learned from running ~700 GPU jobs on Nautilus (MIT license).
Information in Mandarin tones: how much do segmental and pitch (F0) cues each tell us about Mandarin tone? I predicted tones from segmental features, F0 contours, or both, on 1.1M vowels from the AISHELL-1 speech corpus. F0 carried most of the information, and vowel height and rounding added a little more.
- Github: data processing, F0 extraction, and the regression and neural classifiers.
LanguageTesting: A pipeline for analyzing students’ performance in exams, which takes into account item facility, item discrimination, B-index,…
RL Logistics Simulator: we designed a smart logistics simulator using deep Q-learning (DQN) and Dijkstra’s algorithm that can prevent 10% of the traffic jams in Black Friday events via prediction.
Fun
Chongqing Chinese, my home dialect: an IPA illustration of 重庆话, a variety of Southwestern Mandarin, from a native speaker’s recordings: its consonants, vowels, and four tones, with acoustic measurements and a glossed North Wind and the Sun.