Request access to PReMind-Bench
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
PReMind-Bench is released for non-commercial research only under CC BY-NC 4.0. Third-party components keep their own terms (MyVLM: Snap non-commercial license; RAP multi-concept database: non-commercial research; PODS: MIT; DreamBooth: CC BY 4.0; Yo'LLaVA: MIT; Unsplash License photographs). The images include synthetic depictions of hazardous situations (fire, injuries). The test split carries gold labels: do not use it for training or model selection.
Log in or Sign Up to review the conditions and access this dataset content.
PReMind-Bench
PReMind-Bench is the benchmark from PReMind: When Personal Memory Meets Visual Hazards (under review).
An agent decides whether to request personal memory, then answers the request from the scene image and available personal evidence.
Evaluation only. This repository contains the test split only; use it for evaluation, not for training or model selection. The training and validation data are not public: they are shared for non-commercial research after an approved request to the authors. Evaluation code: https://github.com/oyt9306/PReMind.
Image creation. Query images are edited composites created with GPT-Image-2.0 from real scene captures (Mip-NeRF 360 and Zip-NeRF); personal facts are fictitious.
Release scope
| Split | Requests | Scope |
|---|---|---|
test |
1132 | 1020 public main requests plus 112 test-only rows |
conflict |
152 | 140 conflict cases and 12 safe-image controls; a subset of test |
The 140 conflict cases comprise 40 clean main-test requests plus 100 test-only variants. Report conflict resolution on cases only; the safe controls measure false alarms separately. Main on-demand scoring uses 996 released requests. Four Bluey requests and 29 low-quality requests remain excluded; the other 24 public main requests retain the existing top-four scoring exclusions. The paper's main on-demand scores use 999 requests: these 996 plus three of the four requests that cannot be distributed because their image depicts a copyrighted cartoon character.
Each row shows four retrieved memory entries in the frozen display order. The source's online top-four rule and equivalence gate are unchanged. The 12 safe controls are labelled safe / direct answer / no visible acute hazard (SAFE_CONTROL_LABEL_OVERLAY.jsonl records the label changes). They do not require an unprompted emergency adaptation. The reference already describes ordinary assistance and is preserved.
Seen/unseen refers only to the visual hazard: H1 person down is seen; H2--H5 are unseen. A person-down scene with another hazard is counted under the other hazard. There are 63 seen and 77 unseen conflict cases. Controls have null hazard-group/seen fields. Capture scenes and target persons are disjoint between training and test; objects and pets can recur. Personal facts are fictitious.
Configs and loading
retrieved and provenance each have test and conflict splits; images has a media split. There are no train or development examples. test contains conflict, so their row counts must not be added. The column split retains the established per-record test value in Parquet; the dataset split selected when loading identifies the subset. Raw evaluation rows use split: holdout.
Use the local scripts/load_example.py or load the Parquet files directly. scripts/scorer_inputs.py exports the raw scoring inputs without any model. Media paths are relative to this package. The 350 unique images have SHA256 records and per-file license terms.
Main fields: id, scene family_id, request, query image, memory entries with 1-based display indices, risk/action labels, reference answer, expected memory behavior, scorer patterns, construction ID and main_eval/reason. conflict_set identifies the conflict split; conflict_role is conflict_case or safe_control; conflict_variant identifies test-only additions. The control overlay uses G0_closed / G0 / ignore and empty citations/personal support patterns, as for safe image-only main rows. Memory/image entries and retrieval positions are preserved.
All counts and distributions are recomputed in CARD_STATS.json. EXCLUSIONS.json retains the original main exclusions. The authors manually checked every test row.
Sources and terms of use
The dataset (annotations, texts, metadata and our edits) is released under CC BY-NC 4.0 for non-commercial research only. Images from third-party sources are not relicensed and keep their own terms, which also apply:
- MyVLM (Snap Inc.): Snap non-commercial licence (non-commercial, academic use only).
- RAP multi-concept database: non-commercial research use only.
- PODS: MIT.
- DreamBooth: CC BY 4.0; per-image references and licences are listed in
LICENSES.md. - Yo'LLaVA: MIT.
- Unsplash photographs: all photographs of people in the test set were downloaded free from Unsplash and are used under the Unsplash License (free for commercial and non-commercial use). Every other person in a test image is a synthetic person generated for this benchmark. The original Unsplash URLs and photographers were not recorded, so per-image attribution is not available.
- CustomConcept101 and two Wikimedia Commons photographs (memory images only): see
LICENSES.mdfor their terms and attribution. 51 of these files (2 Wikimedia Commons, 49 author-captured CustomConcept101) are under CC BY-SA and are listed one by one, with author and source URL, in the "Per-file licences (CC BY-SA)" table ofLICENSES.md. They are redistributed under their own CC BY-SA licence, with technical changes only (orientation, square padding, resizing); the CC BY-NC 4.0 dataset licence does not apply to them.
Query images are edited composites (GPT-Image-2.0) of scene captures from Mip-NeRF 360 and Zip-NeRF (Barron et al.). They are used, and their edited frames redistributed, with the authors' permission.
Non-commercial sources. Material from MyVLM (Snap non-commercial licence), the RAP multi-concept database and CustomConcept101 files under the Adobe Research License may be used for non-commercial research only. These terms are narrower than CC BY-NC 4.0 and apply to the files that contain them.
LICENSES.md gives the per-source details, attribution and the number of images from each source.
Content warning
The query images include synthetic depictions of hazardous situations, such as fire, smoke, injuries, spills and electrical hazards, that some readers may find distressing. They are synthetic edits made to test whether a personal agent addresses danger first; they do not depict real incidents.
Limitations
Constructed scenes, fictitious personal facts, English requests, model-based judges, household settings and one region's emergency practice limit the benchmark's scope. It evaluates advice rather than executed actions or clinical competence. This local build changes no evaluation result and makes no claim that a particular model passes the benchmark. Test data and golds must not be used for training or model selection. No license or public-release approval is granted by successful validation.
Citation
@misc{premind,
title = {PReMind: When Personal Memory Meets Visual Hazards},
author = {Anonymous},
note = {Under review},
year = {2026}
}
- Downloads last month
- 97