![Donkey_clipart_01.svg.png](/assets/img/Donkey_clipart_01.svg.png?052e3a542c4acd58a0f4e9addfa2965c)
a hidden state is worth a thousand words.
Hi! I am Bokai Xu, I received my bachelor degree at The Chinese University of Hong Kong, Shenzhen. Currently I am working full-time as an engineer at ModelBest Inc (OpenBMB).
My passion is understanding the connection between deep neural networks and biological brain, and the training dynamics in deep neural networks.
About my current research:
At THUNLP@THU aka ModelBest Inc (Jun 2022 - ), I worked on various topics with Large Language Models, I work closely with Dr.Huadong Wang.
- I contributed to language model alignment UltraChat: Enhancing Chat Language Models by Scaling High-quality Instructional Conversations, the dataset were widely used by community, including Falcon, Zephyr and others.
- I contributed to Tool Learning with Foundation Models & BMTools for LLM to use tools, we thoroughly investigated the LLM’s ability to sequentially use tools for completing various tasks.
- I led the one-to-any multimodal dense retrieval Memex: OCR-free Visual Document Embedding Model, enabling VLM to read and comprehend with vision and remember multimodal information and build personal library, in a way like human. We open-sourced enhanced version of contrastive learning training framework, which supports training dense retriver with Vision Language Model.
- I was proud to participate in development of MiniCPM-V-2.6, focusing on enhanced OCR capability and multi-image understanding capability, together with Cuiunbo.
At ICBI@SIAT (Apr 2023 - ), I worked on brain reconstruction infrastructures. I work closely with Dr.Chaoyu Yang and Prof.Pengcheng Zhou.
- I led foundation model pretraining Masked Autoencoders with large-scale microscopic images then aligned pretrained model for trustful brain reconstruction task: Brain Reconstruction by Self Supervised Semantic Stitching of Non-overlapping 3D Microscopic Image .
- My long-term aim is to build scalable pipelines to extract neuronal circuits from microscopic images.
We hope to contribute to the community, we will open-source our progress periodically at RhapsodyAI.