projects
Research I have done, from RL post-training and evaluation to retrieval and open datasets, plus the open-source repositories behind it. Teaching and coursework from Stanford at the bottom.
Research
Open source
- CDial-GPT thu-coai/CDial-GPTLCCC, a 12M-dialogue cleaned Chinese conversation corpus, with pre-trained dialogue models. NLPCC 2020 Best Student Paper.
- CrossWOZ thu-coai/CrossWOZThe first large-scale Chinese cross-domain task-oriented dialogue dataset, with a user simulator and benchmark models. TACL 2020.
- KdConv thu-coai/KdConvA multi-domain knowledge-driven dialogue dataset grounded in knowledge graphs, with benchmark models. ACL 2020.
- ColBERT-Serve stanford-futuredata/colbert-serveMemory-mapped, multi-stage serving for ColBERT with hybrid scoring; 90% less RAM. ECIR 2025.