Access is granted for non-commercial research and evaluation use only. By requesting access you agree not to redistribute this dataset or any derivative data, to cite Dhi Technologies in any publication or output that uses it, and to obtain a separate commercial license via dhi-tech.com before any commercial use. Access requests are reviewed manually by Dhi Technologies.

Causal Predictive Alerting Benchmark

Product: causal-predictive-alerting ("precausal"): predicts an incident seconds before it happens from kinematic trajectories, and records why through ablation replay and a forecast ledger that grades each prediction as fulfilled or falsified.

Paper: Dhi Labs paper 10, Predict, Replay-Ablate, Record: Seeded Model-Ablation Evidence for Predictive Alerting (manuscript in preparation).

Results status. The headline results below are taken from the paper 10 abstract, and the seed sensitivity is taken from its main text. The full set of results, tables and per-condition breakdowns will be posted here when the manuscript is final.

What this dataset is

Seeded, procedurally generated kinematic trajectories (zone entry, convergence, crowd build-up, departure, stationary and out-of-horizon cases), run through the product code. The scenarios in this dataset are synthetic. The public-scene results in paper 10 are not part of these files.

What's in this dataset

battery_200.jsonl (201 lines, about 66 KB): 200 scenario rows plus one trailing summary row, from an earlier 200-scenario battery with the kinds zone_entry, convergence, crowd_buildup and no_incident.
scenario_battery_rows.jsonl (1,996 rows, about 1.58 MB) and scenario_battery_summary.json (about 3.1 KB): a larger battery with eight scenario kinds (zone_entry, convergence, convergence_multi, crowd_buildup, no_incident, departing, stationary, out_of_horizon). Each row carries the counterfactual explanation (would_not_fire_if, still_fires_despite, and per-entity necessity_by_entity).

Scope of these files. The two batteries come from earlier harness runs. The paper 10 results come from the paper's own protocol and its public-scene evaluation. They are not recomputed from these files.

How to load it

import json
from huggingface_hub import hf_hub_download

repo_id = "Dhi-Technologies/causal-predictive-alerting-benchmark"
battery_path = hf_hub_download(repo_id, "battery_200.jsonl", repo_type="dataset")
rows_path = hf_hub_download(repo_id, "scenario_battery_rows.jsonl", repo_type="dataset")
summary_path = hf_hub_download(repo_id, "scenario_battery_summary.json", repo_type="dataset")

battery = [json.loads(line) for line in open(battery_path)]  # last line is a summary, not a scenario
rows = [json.loads(line) for line in open(rows_path)]
summary = json.load(open(summary_path))

print(len(battery), "battery_200 lines")
print(len(rows), "scenario_battery rows")
print(summary["n_scenarios"])

The two files use different schemas. Normalize them before combining.

Results (paper 10 abstract and seed sensitivity)

Seeded synthetic battery, 210 scenarios. Zone-entry precision is 0.667. The designed-negative false-positive rate is 6 of 75.

Public benchmark scenes. Under a protocol written before the results, the unmodified predictor runs on 1,951 annotated pedestrians in five public benchmark scenes (44 minutes of video). Zone-entry precision is 0.568 with recall 0.892.

Equal-budget comparison. At an equal budget of 100 alerts per scene, the kinematic predictor reaches macro precision 0.760, ahead of the rule-based (0.490), motion (0.512) and anomaly (0.378) comparators.

Replay ablation. On real alerts, replay ablation shows that every zone-entry alert depends on its tracked pedestrian.

Vision path. A vision path built from a two-stage or transformer detector, an association-based tracker and published homographies runs at about 17.3 frames per second in a per-frame test on a 16 GB server GPU. On the oblique view, zone-entry recall from detected tracks is 0.744 (0.930 on annotated tracks).

External multi-camera site. A zone-entry policy chosen on the other scenes, and frozen before use, raises zone-entry precision on an external multi-camera site from 0.575 to 0.921.

Seed sensitivity, seeds 0 to 99. The same fixed run, extended from seeds 0 to 9 to seeds 0 to 99 with every other setting unchanged, gives mean precision 0.604 (95 percent bootstrap interval over seeds 0.588 to 0.619) for the kinematic predictor, 0.587 (0.572 to 0.602) for the rule-based comparator, 0.267 (0.255 to 0.280) for the anomaly comparator and 0.412 (0.399 to 0.425) for the motion comparator. In paired differences over seeds, the kinematic predictor exceeds the rule-based comparator by 0.017 (interval 0.007 to 0.027), the anomaly comparator by 0.336 (0.321 to 0.352, higher on all 100 seeds) and the motion comparator by 0.192 (0.177 to 0.206, higher on all 100 seeds).

Method summary

Each fired alert is replayed after deleting or freezing selected track observations (claim scope model-ablation). A separate forecast ledger later resolves each prediction as fulfilled or falsified.

Limitations

The scenarios in this dataset are synthetic. The public-scene results are described in paper 10.
The predictor uses constant-velocity kinematics.

License

This dataset is released under CC BY-NC 4.0 (non-commercial). Access is gated and requires manual approval. It is provided for non-commercial research and evaluation only. Redistribution is not permitted, and any publication or output using it should cite Dhi Technologies. Commercial use requires a separate agreement; contact dhi-tech.com.