heron
Datasets
All datasets matching “heron”heronstegoattack-advbench50
StegoAttack AdvBench-50
Steganographic jailbreak data generated using the StegoAttack pipeline from the paper "Hiding in Plain Sight: A Steganographic Approach to Stealthy LLM Jailbreaks" (Geng et al., 2025).
For experiment results and analysis, see experiment.md.
What is StegoAttack?
StegoAttack is a jailbreak method that uses steganography to hide harmful queries inside benign-looking text. It embeds each word of a harmful query at a fixed position (e.g. the 2nd… See the full description on the dataset page: https://huggingface.co/datasets/heron-ai-security/stegoattack-advbench50.Japanese-Heron-BenchThis dataset is a clarified version of the image, context, and question set included in the Japanese-Heron-Bench for the construction of the Japanese evaluation benchmark suite.
The original dataset refers to turing-motors/Japanese-Heron-Bench.
Link to the original dataset🔗: https://huggingface.co/datasets/turing-motors/Japanese-Heron-Bench
@misc{inoue2024heronbench,
title={Heron-Bench: A Benchmark for Evaluating Vision Language Models in Japanese},
author={Yuichi Inoue and Kento… See the full description on the dataset page: https://huggingface.co/datasets/Silviase/Japanese-Heron-Bench.Japanese-Heron-Bench
Japanese-Heron-Bench
Dataset Description
Japanese-Heron-Bench is a benchmark for evaluating Japanese VLMs (Vision-Language Models). We collected 21 images related to Japan. We then set up three categories for each image: Conversation, Detail, and Complex, and prepared one or two questions for each category. The final evaluation dataset consists of 102 questions. Furthermore, each image is assigned one of seven subcategories: anime, art, culture, food, landscape, landmark… See the full description on the dataset page: https://huggingface.co/datasets/turing-motors/Japanese-Heron-Bench.heroncobalt_heron
wvk-13 duel cover, 60%
Corpus epoch 20 (duel_turns@v4, manifest 61f92695a5aa, 283,828 turns /
9,903 strata), scored under weight_version_key 13. 67,084 turn ids
covering an expected 780 of the 1,300 turns in a duel slice.
Corpus epochs advance without a weight_version_key fork, and every new stratum
dilutes an existing cover. Re-derive against the live manifest before relying on
these ids: the epoch-18 build of this same set now measures 51.4% on live
epoch-20 duels.… See the full description on the dataset page: https://huggingface.co/datasets/iamPi/cobalt_heron.