Transformers | Hunter Heidenreich

Document Processing

Stream accuracy versus relative throughput for Mistral-7B and XGBoost models

LLMs for Insurance Document Automation

LLM applications for insurance document automation using parameter-efficient fine-tuning and analysis of calibration …

Computational Chemistry

SELFIES robustness demonstration

Invalid SMILES Benefit Chemical Language Models: A Study

Skinnider (2024) shows that generating invalid SMILES actually improves chemical language model performance through …

Natural Language Processing

Word vector illustration showing text classification and NLP concepts

Sarcasm Detection with Transformers: A Cautionary Tale

Learn how dataset bias can lead to misleading results in NLP: a sarcasm detection model that actually learned to …

Computational Social Science

Diagram of the Universal Message schema showing fields like ID, Text, Author, and Reply Sets that normalize data across platforms

Look, Don't Tweet: Unified Data Models for Social NLP

PyConversations library and unified data schema for normalizing 300M+ posts across Twitter, Reddit, Facebook, and 4chan.

Computational Social Science

Diagram of the Universal Message schema showing fields like ID, Text, Author, and Reply Sets that normalize data across platforms

PyConversations: Social Media Conversational Analysis

Undergraduate thesis exploring representation learning for social media text and developing tools for cross-platform …

A nonsensical trigger sequence 'WTC theoriesclimate Flat Hubbard Principle' is fed into GPT-2, which then generates Flat Earth conspiracy text

GPT-2 Susceptibility to Universal Adversarial Triggers

Investigation into whether universal adversarial triggers can control both topic and stance of GPT-2's generated text …