Key Skills: Pytorch, Python, AWS

• Finetuning LLMs and developing high-quality test items.

Target: Selecting Better In-Context Learning Demonstrations for Text Classification

Key Skills: TensorFlow, Pandas, Python, NumPy, LaTeX

Models: Flan-PaLM 2, off-the-shelf retriever (fine-tuned on mT5)

• Completed over 100 pages of documentation and 4,000+ lines of code. Prepared a paper for submission.
• Achieved a +2.6% gain on F1 macro scores over an already high baseline that matches or exceeds current benchmarks.
• Proposed constraints for demonstration selection are potentially adaptable to other applications, including ranking.

Target: Generating Questions of Different Styles Controlled with Keywords

Key Skills: PyTorch, PyTorch Lightning, Python, NumPy

Models: T5, mT5, ByT5

• Authored over 3,600 lines of code.
• Demonstrated that an enhanced T5 model with additional tokens, such as emojis, excels in generating keywords together with topics over other models, surpassing spaCy on keyword extraction.
• Generated questions controlled with keywords, topics, and specified length. Determined that using distinct models yields better results for generating questions with different styles.

Target: Generating and Ranking Inquisitive Questions Controlled with Question Types

Key Skills: PyTorch, Fairseq, Pandas, Python, NumPy, LaTeX

Models: RoBERTa, BART

• Code is publicly available on GitHub (5000+ lines). Our paper was accepted for presentation at *SEM 2022.
• Produced diverse questions tailored to specific question types.
• Leveraged a pairwise ranker to select generated questions that matched the quality of human-crafted queries in terms of syntax, semantics, relevancy, and inquisitiveness, as validated by human assessment.