// the find
bojone/bert_in_keras
在Keras下微调Bert的一些例子；some examples of bert in keras
A small collection of Keras examples for fine-tuning BERT on Chinese NLP tasks — sentiment analysis, relation extraction, subject extraction, and NL2SQL. Written in 2019 against TF 1.13 and Keras 2.2.4, it's a historical snapshot, not a maintained library.
The examples are concrete and task-specific rather than toy hello-worlds. The companion blog posts (kexue.fm) apparently go deep on the implementation details. The NL2SQL baseline is a less common example that's harder to find elsewhere. The author links to both the official Google BERT weights and the HIT Chinese WWMBERT weights, which saves time hunting those down.
Python 2.7 + TF 1.13 + Keras 2.2.4 — this hasn't been touched since October 2019 and will not run without a time machine or significant porting effort. There's no package, no tests, no abstraction — just scripts. The entire repo is four Python files. The documentation is in Chinese with external links that may have rotted. Anyone considering this today should look at the author's newer bert4keras project instead.