Hugging Face's logo

Datasets:
xpertsystems
/
ins002-sample

Tasks:
Tabular Classification
Tabular Regression
Time Series Forecasting
Size:
10K<n<100K
Tags:
insurance
claims
auto-insurance
vehicle-accidents
collision
actuarial
License:
Dataset card Data Studio Files
xet
 Community
ins002-sample / README.md
pradeep-xpert's picture
pradeep-xpert
Upload folder using huggingface_hub
f808b59 verified
preview code
|
raw
history blame contribute delete
14.2 kB
metadata 
license: cc-by-nc-4.0
task_categories:
  - tabular-classification
  - tabular-regression
  - time-series-forecasting
tags:
  - insurance
  - claims
  - auto-insurance
  - vehicle-accidents
  - collision
  - actuarial
  - fraud-detection
  - synthetic-data
  - nhtsa
  - kabco
pretty_name: INS-002 — Synthetic Vehicle Accident & Collision Dataset (Sample)
size_categories:
  - 10K<n<100K

INS-002 — Synthetic Vehicle Accident & Collision Dataset (Sample)

XpertSystems.ai Synthetic Data Platform · SKU: INS002-SAMPLE · Version 1.0.0

This is a free preview of the full INS-002 — Synthetic Vehicle Accident & Collision Dataset product. It contains roughly ~9% of the full dataset accidents at identical schema, taxonomy, and benchmark calibration, so you can evaluate fit before licensing the full product.

File Rows (sample) Rows (full) Description
collision_causation_master.csv ~90 ~200 Causation taxonomy reference
accident_period_summary.csv ~1,185 ~720 Period × state × accident-type KPIs
accident_header.csv ~6,524 ~75,000 One row per accident (with outcomes)
accident_transactions.csv ~32,689 ~280,000 Per-transaction event ledger

Dataset Summary

INS-002 simulates the full end-to-end U.S. auto/vehicle accident claims lifecycle with NHTSA-calibrated severity, causation, and outcome modeling, with:

  • 10 collision types: REAR_END, ANGLE, HEAD_ON, SIDESWIPE (same/opposite), FIXED_OBJ, PEDESTRIAN, MOTORCYCLE, ROLLOVER, BACKING — with empirically- calibrated severity matrices (NHTSA CRSS/FARS 2023)
  • 7 vehicle classes: passenger car, SUV/CUV, pickup truck, motorcycle, commercial van, heavy truck, rideshare
  • 6 road classes: interstate, US highway, state route, county road, urban local, rural local
  • KABCO severity scale (K=fatal, A=incapacitating, B=non-incapacitating, C=possible injury, O=PDO) — calibrated to NHTSA empirical distributions
  • MAIS injury severity scoring (Maximum Abbreviated Injury Scale 0-6)
  • 10 causal factors: distraction, speeding, impaired driving, failure to yield, following too close, road condition, mechanical failure, weather/visibility, infrastructure, fatigue
  • 6 treatment types: ED, urgent care, primary care, specialist, PT, imaging
  • 7 weather conditions with state-cluster weather event modeling
  • 4 lighting conditions with calibrated time-of-day distributions
  • Fault attribution: insured/third party/shared with arbitration, litigation, comparative, contributory negligence resolutions
  • Total loss vs repair modeling (ACV vs repair cost threshold)
  • UIM (uninsured motorist) exposure modeling
  • Fraud injection: staged collisions, inflated repairs, phantom injuries, weather-period fraud uplift
  • Weather surge events with FNOL volume amplification (3.6× baseline)
  • 12 U.S. states with state-cluster regulatory variation

Calibrated Benchmark Targets

The full product is calibrated to 12 benchmark validation tests drawn from authoritative sources (NHTSA CRSS/FARS 2023 data, IIHS, Insurance Information Institute, Coalition Against Insurance Fraud, NAIC, McKinsey, WCRI, state DOI/insurance department data):

Test Target Observed Verdict
total_loss_rate 0.2100 0.1758 ~ MARGINAL
bodily_injury_rate 0.3000 0.2719 ✓ PASS
fatality_rate 0.0105 0.0150 ✗ FAIL
avg_repair_cost_usd 4850.0 4952.3 ✓ PASS
avg_mais_score_injury_accidents 1.8400 1.6921 ✓ PASS
fault_dispute_rate 0.1800 0.1810 ✓ PASS
uim_identification_rate 0.1400 0.1409 ✓ PASS
weather_surge_ratio 3.6000 3.7487 ✓ PASS
fraud_referral_rate 0.1100 0.1188 ✓ PASS
avg_cycle_time_days_pdo 18.5000 17.9300 ✓ PASS
avg_cycle_time_days_bi 52.0000 51.8400 ✓ PASS
reserve_adequacy_rate 0.9300 0.9257 ✓ PASS

Note: fatality_rate and total_loss_rate are rare-event benchmarks that require larger sample sizes to fully converge to target. At the sample's ~6,500 accidents, observed fatality_rate has ±30% variance relative to the 0.0105 target. The full product (75,000 accidents) demonstrates these benchmarks at full statistical power.

Schema Highlights

accident_header.csv (primary file)

Column Type Description
accident_id string Unique accident identifier
claim_id string Associated claim FK
policy_number string Policy identifier
carrier_id string Carrier entity
accident_date date Date of accident
report_date date Date reported (FNOL)
close_date date Date closed (nullable)
cycle_time_days int Days from FNOL to close
state string U.S. state code
collision_type string 1 of 10 collision types
vehicle_class string 1 of 7 vehicle classes
road_class string 1 of 6 road classes
kabco_severity string K/A/B/C/O — NHTSA severity scale
mais_score int MAIS injury severity (0-6)
weather_condition string 1 of 7 weather conditions
lighting_condition string 1 of 4 lighting conditions
primary_causal_factor string 1 of 10 causal factors
secondary_causal_factor string Secondary causal factor (nullable)
speed_limit_mph int Posted speed limit
estimated_speed_mph int Estimated speed at impact
n_vehicles_involved int Vehicle count
n_persons_involved int Person count
n_injuries int Injury count
n_fatalities int Fatality count
fatality_flag int Boolean — fatality occurred
bodily_injury_flag int Boolean — BI claim
pdo_only_flag int Boolean — property damage only
repair_cost_usd float Repair cost
acv_usd float Actual cash value
total_loss_flag int Boolean — repair > total-loss threshold
medical_cost_usd float Medical expenses
treatment_type string Primary treatment category
weather_event_id string FK to weather event (nullable)
weather_event_flag int Boolean — during weather surge
fault_party string INSURED / THIRD_PARTY / SHARED
fault_dispute_flag int Boolean — fault disputed
fault_resolution string agreed / arbitrated / litigated / comparative
uim_exposure_flag int Boolean — UIM exposure identified
denial_flag int Boolean — claim denied
denial_reason string Categorical denial reason
reopen_flag int Boolean — reopened
fraud_referral_flag int Boolean — referred to SIU
staged_collision_flag int Boolean — fraud subtype
inflated_repair_flag int Boolean — fraud subtype
phantom_injury_flag int Boolean — fraud subtype

accident_transactions.csv (event ledger)

Column Type Description
transaction_id string Unique transaction ID
accident_id string Parent accident FK
claim_id string Parent claim FK
txn_date date Transaction date
txn_type string RESERVE / PAYMENT / RECOVERY / ADJUSTMENT
txn_amount float Transaction amount (signed)
reserve_balance float Reserve balance after txn
paid_balance float Paid balance after txn
adjuster_note string Free-text note (synthetic, anonymized)
description string Transaction description

collision_causation_master.csv

Reference master with synthetic causation taxonomy descriptions for each collision_type × primary_causal_factor combination.

accident_period_summary.csv

Aggregate KPIs by (period, state, collision_type): accident counts, average severity, fatality counts, fault dispute rates, total loss rates.

Suggested Use Cases

  • Training accident severity classification models (KABCO/MAIS)
  • Fault attribution prediction — predict at-fault party from accident features
  • Total loss vs repair decisioning — predict total loss likelihood
  • Bodily injury claim flagging — distinguish PDO from BI at FNOL
  • Fatality risk scoring by collision type, vehicle class, road type
  • Fraud detection — staged collision, phantom injury, inflated repair
  • Weather event surge forecasting for catastrophe reinsurance
  • UIM (uninsured motorist) exposure prediction
  • Cycle time forecasting (PDO vs BI claims have different lifecycles)
  • Reserve adequacy modeling for auto claims
  • Claim denial prediction with categorical denial reasons
  • Telematics + accident severity correlation (when joined with telematics)
  • Litigation prediction by fault resolution mode
  • Pedestrian/motorcycle vulnerability modeling
  • Distracted driving / impaired driving causal modeling
  • Insurtech actuarial model training with NHTSA-calibrated synthetic data

Loading the Data 

import pandas as pd

headers      = pd.read_csv("accident_header.csv",
                           parse_dates=["accident_date", "report_date", "close_date"])
transactions = pd.read_csv("accident_transactions.csv", parse_dates=["txn_date"])
causation    = pd.read_csv("collision_causation_master.csv")
summary      = pd.read_csv("accident_period_summary.csv")

# Multi-class collision-type classification target
y_collision = headers["collision_type"]

# Binary total-loss prediction target
y_total_loss = headers["total_loss_flag"]

# Binary bodily-injury target
y_bi = headers["bodily_injury_flag"]

# Multi-class KABCO severity target
y_severity = headers["kabco_severity"]

# Binary fraud-referral target
y_fraud = headers["fraud_referral_flag"]

# Regression: repair cost prediction
y_repair_cost = headers["repair_cost_usd"]

# Multi-class fault attribution target
y_fault = headers["fault_party"]

# Aggregate per-accident payment trajectories (sequence modeling)
payment_sequences = transactions[transactions["txn_type"] == "PAYMENT"] \
    .groupby("accident_id")["txn_amount"].apply(list)

License

This sample is released under CC-BY-NC-4.0 (free for non-commercial research and evaluation). The full production dataset is licensed commercially — contact XpertSystems.ai for licensing terms.

Full Product

The full INS-002 dataset includes ~355,000 rows across all four files, with 12 calibrated benchmark validation tests drawn from authoritative auto/vehicle accident data sources (NHTSA CRSS/FARS 2023, IIHS, Insurance Information Institute, Coalition Against Insurance Fraud, NAIC, state DOI data).

📧 pradeep@xpertsystems.ai 🌐 https://xpertsystems.ai

Citation 

@dataset{xpertsystems_ins002_sample_2026,
  title  = {INS-002: Synthetic Vehicle Accident & Collision Dataset (Sample)},
  author = {XpertSystems.ai},
  year   = {2026},
  url    = {https://huggingface.co/datasets/xpertsystems/ins002-sample}
}

Generation Details

  • Generator version : 1.0.0
  • Random seed : 42
  • Generated : 2026-05-16 19:53:01 UTC
  • Architecture : Benchmark-first — each metric → one named parameter
  • Overall benchmark : 91.8 / 100 (grade A-)