● Extended vocabularies of Open-Llama, Bloom, StarCoder to include Vietnamese through continual pre-training. ● Integrated BLOOM into ExLlama inference framework (4-bit Pytorch), resulting in threefold reduction in inference time. ● Constructed a Vietnamese evaluation benchmark containing national university entrance mock tests, guaranteeing contamination-free, coverage and difficulty.

更多

● Analyzed the temporal reasoning capabilities of large language models across various NLP tasks. ● Implemented continual training of T5 by masking temporal keywords, employing a generator-discriminator mechanism to solve TRACIE dataset (implicit event temporal reasoning using common-sense knowledge). ● Conducted experiments with BLOOMZ-176b and ChatGPT using diverse prompting techniques to tackle temporal reasoning. ● Successfully addressed TRACIE dataset using ChatGPT in both chain-of-thought few-shot and task-explanation prompting settings. ● Customized and fine-tuned speech processing models from NeMo toolkits for the analysis of Korean interviews.

更多

● Improved relevant responses suggestion model by updating dataset with database’s multiple instances, adding a new API request that query responses with specific topics. ● Added new features concerning temporal information of students’ performance, thus improving level assessment models by 3% F1-score. ● Fine tuned and monitored DialoGPT for the domain tasks with corresponding datasets.

更多

● Build an interactive and customizable data mathematical presentation tool using REDOM and Fabricjs. The tool supports visualizing 1D and 2D datasets in multiple plot types and accessing their statistical figures. ● Improve the file management system by implementing and standardizing file export functions supporting a wide range of input data types and file formats.

更多

Extended a machine learning fairness and robustness improvement training technique with a larger dataset and multiple-sensitive-feature poisoning attacks (group-targeted label flipping).

更多