Datasets:
Download README.md from WZDavid/JARVIS-Bench: direct link, hf CLI and curl.
- Browser
- Download file 5.07 kB
-
https://huggingface.co/datasets/WZDavid/JARVIS-Bench/resolve/main/README.md
- Command line
-
hf download hf://datasets/WZDavid/JARVIS-Bench/README.md
-
curl -L -o README.md https://huggingface.co/datasets/WZDavid/JARVIS-Bench/resolve/main/README.md
license: cc-by-nc-4.0
pretty_name: JARVIS-Bench
size_categories:
- 10K<n<100K
tags:
- benchmark
- personalization
- long-horizon-memory
- synthetic-data
configs:
- config_name: events
default: true
data_files:
- split: full
path: data/events.parquet
- config_name: days
data_files:
- split: full
path: data/days.parquet
- config_name: profiles
data_files:
- split: full
path: data/profiles.parquet
JARVIS-Bench
Project homepage: https://agentmemoryworld.github.io/jarvis-bench/
JARVIS-Bench studies personal intelligence agents using longitudinal daily-life traces, including real-log-derived anchors and synthetic trajectory expansions. The intended research tasks are implicit preference reasoning, proactive help prediction, emotion-aware modeling and cross-domain transfer.
The dataset includes longitudinal event records, day-level metadata and user profiles.
Contents and statistics
| Item | Count |
|---|---|
| Users with assembled trajectories | 42 |
| User-days | 1,495 |
| Events | 18,529 |
Events marked is_real_log=true |
9,437 |
Events marked is_real_log=false |
9,092 |
Events marked has_image=true in source |
2,066 |
| Days per assembled trajectory | 28–53 |
Note: Descriptions corresponding to the images are included in the textual records; image files are not distributed in this repository.
Load
from datasets import load_dataset
events = load_dataset("WZDavid/JARVIS-Bench", "events", split="full")
days = load_dataset("WZDavid/JARVIS-Bench", "days", split="full")
profiles = load_dataset("WZDavid/JARVIS-Bench", "profiles", split="full")
For local use, pass data_files="data/events.parquet" to load_dataset("parquet", ...).
Pin a Hub commit using revision=... when reporting results. full is the exported corpus,
not a training or held-out test assignment. No random split has been introduced.
Schema and reconstruction
events:user_id,day_key,event_key,event_id,day_index,seq,domain,time_bucket,is_real_log,has_image,summary_text,description_long,image_available,event_json.days:user_id,day_key,day_index,day_metadata_json(all day fields except events).profiles:user_id,profile_json,trajectory_metadata_json(all trajectory fields except days).
JSON-suffixed columns contain UTF-8 JSON strings; decode with json.loads.
event_json preserves every source event field, including payload, provenance, clues and image references.
Join days and events on (user_id, day_key) and use event_key to reconstruct the original event map.
Sort by day_index and seq for chronological access. is_real_log records source provenance;
it does not certify that expanded prose is a verbatim participant statement.
Evaluation boundaries
Profiles contain survey-derived insights and may reveal evaluation targets. Only expose them under an explicitly permitted protocol (e.g. a documented oracle condition). Do not pass future events, labels, rationales or implicit-clue annotations to a model unless the task permits them. Synthetic-event annotations are not substitutes for participant ground truth. This package does not provide a standalone 100-question survey instrument, exact paper task manifests, executable evaluation harness or verified task-specific test splits. The full corpus contains 18,529 events. The paper reports 18,471 events for its evaluation set; use the paper's task-specific protocol when reproducing evaluation results.
Collection, provenance and responsible use
The dataset was collected through a voluntary UIC IRB-approved study with informed consent,
70 enrolled participants and 42 retained participants. It describes participant-authored logs,
personas and survey responses, followed by image grounding and synthetic trajectory expansion.
Collection, annotation and processing details are documented in croissant.json.
The cohort is small, author-reported demographics are US-concentrated, and synthetic content can introduce generator biases. Demographic language shares are not measured text-language shares. Personal routines, mood, preferences and demographics can be sensitive even with pseudonymous IDs. The collection protocol includes informed consent, privacy processing and participant fidelity confirmation. This export preserves the supplied released text. The dataset is intended for research evaluation; results do not establish deployment safety or clinical validity.
License and citation
The dataset is licensed under CC BY-NC 4.0. See LICENSE for the full terms.
Weizhi Zhang et al. (2026). JARVIS-Bench: Benchmarking Personal Intelligence Agents on Long-Horizon Real-User Daily Traces. Version 1.0.0. https://huggingface.co/datasets/WZDavid/JARVIS-Bench
Full creator names are in croissant.json. The dataset is hosted on Hugging Face.
Report issues through the dataset repository's community discussion; no fixed maintenance cadence is promised.