New Python package GRADIEND enables gradient-based feature learning in LLMs

Researchers have introduced GRADIEND, an open-source Python package designed for gradient-based feature learning in language models. This package offers a comprehensive workflow that includes data creation, model training, evaluation, and persistent model rewriting through controlled weight updates. The GRADIEND method focuses on learning feature directions from factual-counterfactual MLM and CLM gradients, and the package facilitates multi-feature comparisons, demonstrated with examples in English pronoun resolution and semantic sentiment analysis. AI

IMPACT Enables researchers to more easily explore and manipulate feature representations within language models.

RANK_REASON The cluster describes a new open-source package for a specific research method in NLP, detailed in an arXiv paper. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

• arXiv
• GRADIEND
• Hugging Face
• Jonathan Drechsel
• language models
• Python

• paper
• infra

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →