ary
Datasets
All datasets matching “ary”locos-results
LOCOS Results
Consolidated results for the LOCOS project: retrieval-head detection, head-ablation
experiments, and downstream long-context evaluations. This single repository
replaces the earlier split across aryopg/decore-results (heads/ablation) and
aryopg/locos_downstream_results (downstream evals).
Paper: Logit-Contribution Scoring Identifies Non-Literal Retrieval Heads
Project Page: https://aryopg.com/locos/
Code: github.com/aryopg/locos. The deploy job scripts there… See the full description on the dataset page: https://huggingface.co/datasets/aryopg/locos-results.CatVision
CatVision: Human–Cat Vision Frame Pairs
Official dataset for the paper:
Purrturbed but Stable: Human-Cat Invariant Representations Across CNNs, ViTs and Self-Supervised ViTsArya Shah et al. · arXiv:2511.02404
This dataset contains 346,400 paired video frames rendered under human vision and simulated cat vision optics.
It was used to benchmark cross-species representational alignment across CNNs, supervised ViTs, windowed transformers,
and self-supervised ViTs (DINO) using CKA and… See the full description on the dataset page: https://huggingface.co/datasets/aryashah00/CatVision.xBDGaslight-Gatekeep-V1-V3
Gaslight, Gatekeep, V1–V3
Early Visual Cortex Alignment Shields Vision-Language Models from Sycophantic Manipulation
Dataset Summary
This dataset accompanies the paper "Gaslight, Gatekeep, V1–V3: Early Visual Cortex Alignment Shields Vision-Language Models from Sycophantic Manipulation". It contains two components:
Gaslighting Benchmark (gaslighting_prompts_v2.json / Parquet): 6,400 structured two-turn adversarial prompts designed to test sycophantic… See the full description on the dataset page: https://huggingface.co/datasets/aryashah00/Gaslight-Gatekeep-V1-V3.04-wikineuralGeoCLIP-data
