DEFault++ benchmark v1
收藏DataCite Commons2026-05-04 更新2026-05-07 收录
下载链接：
https://zenodo.org/doi/10.5281/zenodo.20018623
下载链接
链接失效反馈官方服务：
资源简介：
DEFault++ benchmark v1: Hierarchical fault diagnosis of fine-tuned HuggingFace transformers
Contents- encoder_merged.csv (6,042 rows × 2,121 cols): trainer-ready feature vectors for 4 encoder models × 6 GLUE tasks (BERT, DistilBERT, RoBERTa, DistilRoBERTa on MRPC, QNLI, QQP, RTE, SST-2, STS-B).- decoder_merged.csv (2,535 rows × 3,130 cols): same shape for 4 decoder models × 5 LM tasks (GPT-2, DistilGPT-2, GPT-Neo-125M, OPT-125M on Lambada, OpenWebText, PTB, WikiText).- 35 per-task source CSVs and feature dictionary for full reproducibility.- Pre-computed deduplication, NaN handling, log-transforms, and CV filtering applied; downstream layer aggregation runs at training time via FeatureProcessor.
Companion code: https://github.com/SigmaJahan/DEFaultplusplus-Transformer-Debugging
Built from fine-tuning runs of 11,251 encoder + 4,282 decoder configurations covering ~13 fault categories and ~44 root causes (see feature_dictionary.csv for the complete catalog).
Citation: see the DOI on this page.
提供机构：
Zenodo
创建时间：
2026-05-04
