Missed the latest GitHub check; these facts may be older.
Overview
CLUE is a benchmark for Chinese NLP. It bundles classification and reading-comprehension datasets (AFQMC, TNEWS', IFLYTEK', OCNLI/CMNLI, CLUEWSC2020, CSL, CMRC2018, DRCD, ChID, C3) with a diagnostics set. It also provides TensorFlow and PyTorch baseline scripts, the PyCLUE toolkit, a 14 GB pre-training corpus and a leaderboard at cluebenchmarks.com.
Who it is for: NLP researchers evaluating Chinese language models
Strengths
Ten Chinese tasks with documented split sizes, formats and per-task download links
Baseline shell scripts fetch model and data, and write submission JSON
Hard-sample filtering by k-fold cross-validation is documented for the v1 datasets
Includes a 14 GB pre-training corpus and a CMNLI-based diagnostics set
Weaknesses
Leaderboard tables cover BERT-era encoders, not recent LLMs
Baselines require TensorFlow 1.12, CUDA 9.0 and cuDNN 7.0
No license detected, so reuse terms for code and data are unclear
Some corpus files are on Baidu Pan or by emailed request
What it needs
GPU optional
Needs TensorFlow 1.12, CUDA 9.0, cuDNN 7.0, PyTorch