Datasets:
Formats:
parquet
Size:
10K - 100K
ArXiv:
Tags:
robustness
adversarial-evaluation
prompt-injection
decision-models
multimodal
arxiv:2610.19118
Commit History
7 days ago
-
Add five decision-model configurations and October 26 results update
Verified signaturePublish complete curated results for d3-turbo-27B, Perplexity Decider v1.2-27B / Flash v1.2-4B, OneJev1.6-27B and local StartLux Instant-27B. Include sizes, checkpoint pins, settings, ASR/accuracy/coverage and group breakdowns. Distinguish v1.3.0 additions from historical v1.2.0 results; preserve all data and prior metrics.f2969d6
13 days ago
-
Release v1.3.0: five full pools and portable GUI identifiers
Verified signaturePublish 19,190 full examples including the curated subset; retain curated defaults and historical model metrics. Update 68 GUI candidate identities and disclose six affected curated inputs. No full model results.f4dcec8
18 days ago
-
Add 26 model configurations and JEV image results with versioned settings
Verified signature92edfdd
22 days ago
-
Report targeted ASR for all prompt injection suites
Verified signatureRescore archived predictions against explicit injection targets. Preserve candidate membership, GT, coverage, denominators and native scoring. Keep untargeted ASR for text adversarial and image perturbation. Archive flip metrics, publish attack-target sidecar and reusable scoring helper, and clarify original flip-based selection. No inference or raw dataset changes.2bd4c63 -
Evaluate OpenAI Decisions on all three public image attack sets
Verified signatureAdd gpt-6-luna Decisions results on all 543 frozen image attack candidates. Preserve all prior results, records and images. Include ASR, refusal counts, native image input settings and per-group breakdowns.05abce6 -
Complete image perturbation coverage on sixteen decision models
Verified signatureAdd eight post-selection model evaluations on all 200 frozen image-perturbation pairs. Preserve all prior metrics, records, image bytes and the v1.2.0 data revision. Include ASR, counts, native model bindings and per-group breakdowns.6d27a00
23 days ago
-
Add seven decision-model evaluations on the frozen curated release
Verified signatureAdd H2O-Lightning-4B, Liquid d1-3B, Perplexity Decider v1.1-27B, InnerJev-27B, GLiNER2.5-Decide-340M, CLM-8B and Laya multilingual. Include ASR counts, coverage/exclusions, clean accuracy, per-source results and pinned settings. Preserve previous metrics and all 1,058 records/images. Data version/tag remains v1.2.0.3c71ea6 -
Show clean and attacked images first in Dataset Viewer
Verified signatureReorder display columns across five configurations. Preserve all 1,058 records, embedded PNG bytes, native JSONL and evaluation results.0a2bdd9 -
Prepare public release and note forthcoming paper
Verified signaturee39259c -
Release v1.2.0: add 13 human-reviewed image coverage supplements
Verified signaturePreserve existing records and IDs. Image perturbation: 200 candidates / 94 originals, including 127 Common and 73 optimization examples. All five sets: 1,058 candidates / 441 originals. Update ASR, composition and dataset citation. Original five severity levels only; no low-strength exploratory data.8b6bac5 -
v1.1.0: remove overperturbed Common images consistently across scoring items
Verified signaturea8ad58a -
Simplify homepage quick start and move usage details to documentation
Verified signaturec18c228 -
Release five curated DecisionRobust configurations v1.0.0
Verified signature600e9fb