philiphoward/temporal-nora

Temporal NoRA HumanGold-190

Temporal NoRA adds judgments of when it is reasonable to start an action to the HumanGold-190 subset of NoRA. It covers five time bands, from the next 10 seconds to three hours.

Contents	Count
Clips	190
Actions	4,913
Action-band judgments	15,433
Positive judgments	11,397
Negative judgments	4,036

The dataset contains text annotations and identifiers. Raw video, audio, and frames are not included. All records belong to the held-out test split.

Load the data

After uploading this folder's contents to a Hugging Face dataset repository, replace YOUR_USERNAME/YOUR_DATASET with the repository ID:

from datasets import load_dataset

dataset = load_dataset("YOUR_USERNAME/YOUR_DATASET", split="test")

To read the local JSONL file with the Python standard library:

import json

with open("data/test.jsonl", encoding="utf-8") as file:
    records = [json.loads(line) for line in file if line.strip()]

Data format

data/test.jsonl contains one JSON object per clip. Each record includes:

Field	Description
clip_id	Identifier linking the record to the source clip.
facts	Scene facts, each with a fact_id and text.
band_applicability	Whether each time band can be assessed, with reasons for exclusions.
canonical_actions	Actions, conditions, and judgments by time band.
annotation_summary	Per-clip annotation counts.
split	Always test.

Within each action, horizon_judgments maps band IDs to judgments. reasonable is the final Boolean label. yes_votes retains the original positive vote count out of three. adjudicated identifies decisions resolved through adjudication. Positive judgments include a reason and supporting fact_ids or foundation. Additional fields retain source references, versions, and export timestamps.

Time bands

Band	Time to start the action
H10	0 to 10 seconds, inclusive
H60	More than 10 seconds, up to 1 minute
H30M	More than 1 minute, up to 30 minutes
H1H	More than 30 minutes, up to 1 hour
H3H	More than 1 hour, up to 3 hours

Bands describe when to initiate an action, rather than its duration. Judge each band separately using the scene facts and action conditions. Assume urgent earlier actions have been addressed unless the action explicitly concerns an unresolved situation. Excluded bands are missing assessments, not negative labels.

Use and attribution

This dataset supports evaluation of temporal action reasonableness. Preserve the held-out test split. Labels reflect judgments in the recorded scenes and do not establish universal rules.

Credit the Temporal NoRA project group, COMP3988_F12_03, and the parent NoRA collection, MINTLABJHUANU/NoRA. Use and redistribution remain subject to the parent dataset's licence and access conditions. This package grants no additional rights.

Release: temporal-nora-humangold190-revised-v1. Parent revision: c85b874f6f36bd7557750d7fa678bae5631c6143.