SafetyJev Unsafe600
Training-ready data for Classifier and Predictor Judge, constructed from 600 unsafe robot rollouts across Cabinet, Clutter, Dusty, Jar, Lid and Stack. The approved episode split is 480 train / 60 validation / 60 test, with whole task-groups kept together. Labels, action windows and train-only normalization are already prepared. No raw conversion or semantic reconstruction is needed.
Choose a training package
| Package | Eligible windows | Input and task |
|---|---|---|
| Classifier | 5,765,453 | Current images + safety question → current violation |
| Predictor Judge | 727,025 | Previous executed actions/images/robot states + current state/question + next8actions → violation prediction |
Both packages use the same episode resources. Download the complete dataset;
point data.package to unsafe600/classifier or unsafe600/predictor_judge.
Decoded caches are optional: set data.frame_cache=null for direct reading.
Data interfaces have passed acceptance; publication of compatible SafetyJev
training code is pending review. Use the reviewed code commit from the handoff.
Browse the data
README.md
classifier/ Classifier indexes, rules and package metadata
train/part-00000.jsonl Numbered storage parts, read automatically
validation/ Natural validation distribution
test/ Natural test distribution
excluded/ Exclusion reasons
predictor_judge/ Same layout for future-violation prediction
episodes/ Shared by both models
cabinet/gr00t_cabinet_task_0002_location_seed0/
trajectory.hdf5 Original JPEG overview/wrist bytes + offsets only
record.json State/action timeline and preserved source evidence
annotation.json Semantic supervision (not model input)
image_manifest.json Per-image integrity hashes
clutter/ dusty/ jar/ lid/ stack/
metadata/
episodes.json Episode ID ↔ readable directory, policy and split
splits/ Task-group assignment and development provenance
composition.json Label distributions and exclusions
validation.json Data acceptance scope
release.json Data/code release status
Folder names describe task family, policy, task, condition and seed. Original episode IDs remain unchanged inside records and indexes. Storage parts and episode folders are independent of the machines used to build the dataset.
Download and use
After gated access is granted, authenticate your Hugging Face client and download the exact dataset revision from the collaborator handoff (or the commit shown in this repository's History):
from huggingface_hub import snapshot_download
snapshot_download("IDEAS-Lab-Northwestern/SafetyJev-unsafe600",
repo_type="dataset", revision="<pinned-dataset-revision>",
local_dir="datasets/unsafe600")
Environment, training and evaluation commands live in the SafetyJev Classifier guide and Predictor Judge guide.
Interpretation
Train sampling can target60% Yes without changing labels; validation/test keep natural distributions. Windows overlap and are not independent incidents. Most Judge Yes windows already start in violation; current-safe→future-violation windows must be reported separately. See composition. All source episodes were selected as unsafe: this cohort alone cannot establish safe-episode false-alarm rates. Known development exposure and historical split differences are retained; test is not claimed to be untouched.
The separate raw archive is optional and is not needed for training. Simulator runtime/assets are not included. Original source evidence is preserved in records for traceability; loaders expose only the documented model input fields.
- Downloads last month
- 192