THU-KEG/SAEVerbalizer-Data

SAEVerbalizer Data

This repository hosts the datasets released for SAEVerbalizer: Generating Explanations for Sparse Autoencoder Features via Representation Verbalization.

The initial release contains the three evaluation sets used for the THU-KEG/SAEVerbalizer-27B verbalizer. Training data and data for other components may be added in later release stages.

The inference and evaluation code is available at THU-KEG/SAEVerbalizer.

Repository structure

evaluation/
├── global_train_standard_1000.json
├── low_index_gold_200.json
└── global_gold_1000.json

Test sets

Test set	Abbreviation	File	Examples
Global Train-Standard	GTS	evaluation/global_train_standard_1000.json	1,000
Low-Index Gold	LIG	evaluation/low_index_gold_200.json	200
Global Gold	GG	evaluation/global_gold_1000.json	1,000

GTS contains globally sampled features satisfying the training qualification standard. LIG and GG satisfy the stricter gold qualification standard; LIG is sampled from the low-index region, while GG is sampled globally. The three sets are mutually disjoint.

Each JSON file is a list of records with the following fields:

• id: SAE feature index.
• concept: reference natural-language feature explanation from Neuronpedia.

The feature IDs refer to the layer-16 width-262k l0_medium SAE from google/gemma-scope-2-27b-it.

Citation

@article{meng2026saeverbalizer,
  title   = {SAEVerbalizer: Generating Explanations for Sparse Autoencoder Features via Representation Verbalization},
  author  = {Meng, Weihan and Guo, Hongzhu and Jing, Yi and Liu, Dewen and Yao, Zijun and Wang, Xiaozhi and Hou, Lei and Li, Juanzi},
  journal = {arXiv preprint arXiv:2608.13538},
  year    = {2026}
}