Alexander Vishnevskiy Email & Phone Number

• Developed KTO fine-tuning technique to enhance question generation for technical documents, resulting in a 10% improvement in top-5 recall.• Enhanced chatbot performance by implementing ReAct prompting techniques, leading to a 25% increase in multi-hop question-answering accuracy.• Applied advanced filtering techniques to train models locally, reducing error rates by 40%.

San Jose, California, United States

Explored Direct Preference Optimization (DPO), an innovative offline algorithm for aligning language models with human preferences.Successfully reproduced the original DPO paper, validating its methodologies and results.Conducted in-depth analysis of the fine-tuning process, with a focus on model forgetting phenomena.Investigated the impact of the beta hyperparameter on model performance, exploring the trade-offs between alignment and base model deviation.Quantified improvements of DPO over standard Supervised Fine-Tuning (SFT) approaches, demonstrating enhanced alignment with human preferences.Contributed to the AI community by fine-tuning and releasing an improved version of the BTLM-3B-8K model using DPO, now available on Hugging Face: https://huggingface.co/cerebras/btlm-3b-8k-chat.

• Developed a scalable PySpark data pipeline to aggregate and preprocess data from multiple sources for training an item-to-item matching re-ranker model.• Implemented advanced sample collection strategy using ColBERT and BERT models to generate high-quality training data for an item matching system.• Built a BERT-based algorithm to match similar articles from item descriptions, improving precision by 0.05 and utilizing Precision-Recall metrics for future training predictions.

• Engineered an advanced credit scoring system using LSTM networks and multi-task learning.• Architected and deployed a high-performance distributed data pipeline for 500,000+ users, reducing scoring time by 80%.• Designed a feature engineering framework extracting 200+ nuanced features, improving ROC-AUC by 2.5%.