Datasetpublic

farshelina/ACL-Bench-Task-Transfer

ACL Bench ACL Bench is a reasoning benchmark developed to evaluate structural task transfer after rule-constrained fine-tuning of large language models. It abstracts structural operations from Dou Dizhu into three reasoning categories: Hierarchy Logic Pattern Recognition Counter Logic Dataset Size 300 multiple-choice questions: 100 Hierarchy Logic 100 Pattern Recognition 100 Counter Logic Project Paper: Rule-Constrained Fine-Tuning and Task… See the full description on the dataset page: https://huggingface.co/datasets/farshelina/ACL-Bench-Task-Transfer.

sourceHugging Faceupdated 18d agoView on Hugging Face

0likes62downloads

Dataset Card

ACL Bench

ACL Bench is a reasoning benchmark developed to evaluate structural task transfer after rule-constrained fine-tuning of large language models.

It abstracts structural operations from Dou Dizhu into three reasoning categories:

—Hierarchy Logic
—Pattern Recognition
—Counter Logic

Dataset Size

300 multiple-choice questions:

—100 Hierarchy Logic
—100 Pattern Recognition
—100 Counter Logic

Project

Paper: Rule-Constrained Fine-Tuning and Task Transfer in Large Language Models: Behavioral Evaluation and Causal Analysis of Attention Heads

Live Demo

Interactive Demo: https://rule-constrained-task-transfer.streamlit.app/

GitHub Repository: https://github.com/farshelina/rule-constrained-task-transfer

Paper

Coming soon.