README.md

---
license: cc-by-nc-4.0
task_categories:
  - tabular-classification
  - tabular-regression
language:
  - en
tags:
  - synthetic
  - vr-training
  - immersive-training
  - operator-competency
  - emergency-response
  - oil-and-gas
  - safety-training
  - simulator-fidelity
  - human-performance
  - opito
  - ipieca
  - iadc-well-control
  - nfpa-1006
  - ccps
  - nebosh
  - api-rp-755
  - uk-hse-ohra
  - dnv-rp-a203
  - process-safety
  - training-analytics
pretty_name: "OIL-046 — Synthetic Training Simulation Dataset (Sample)"
size_categories:
  - 100K<n<1M
---

# OIL-046 — Synthetic Training Simulation Dataset (Sample)

A schema-identical preview of **OIL-046**, the XpertSystems.ai synthetic
**VR-based operator training simulation** dataset for upstream + offshore +
refinery oil & gas operations. The full product covers ~350,000 trainees
× ~8,500 facilities × ~85 million sessions across a 5-year horizon. This
sample is HF-sized (500 trainees × 30 facilities × 2,500 sessions ×
15,000 VR movements) covering all 13 product tables.

> **Built by** XpertSystems.ai — Synthetic Data Platform
> **Contact** [pradeep@xpertsystems.ai](mailto:pradeep@xpertsystems.ai) · [xpertsystems.ai](https://xpertsystems.ai)
> **License** CC-BY-NC-4.0 (sample); commercial license available for the full product.

---

## What OIL-046 does that nothing else in the catalog does

OIL-046 is the catalog's **first VR / immersive training simulation** SKU.
Where OIL-035 (Safety / HSE) models incidents *after the fact* and OIL-045
(Workforce) models scheduling and fatigue, OIL-046 models the **training
data** that determines whether operators are competent to handle those
incidents when they occur. This is the substrate underneath every other
safety-related dataset in the catalog.

This is the substrate **VR training platform vendors, operator competency
analytics teams, simulator fidelity researchers, OPITO/IADC training
auditors, and human-performance modelers** have been waiting for: a
coherent, joinable dataset where VR sessions, equipment interactions,
alarm acknowledgments, communication chains, evacuations, safety
violations, fatigue, and incident progression all share session_id and
trainee_id for cross-modal training analytics.

| Buyer Persona | Use Case |
|---|---|
| VR Training Platform Vendor | Simulator fidelity validation, scenario completion analytics |
| Operator Competency Analytics | Pass/fail prediction, retraining recommendation models |
| Simulator Fidelity Researcher | VR realism scoring, hardware profile impact |
| OPITO/IADC Training Auditor | Compliance reporting + competency benchmarking |
| Human-Performance Modeler | Fatigue-in-training × decision quality × stress correlation |
| HSE Training Director | Drill effectiveness + violation pattern detection |
| Insurance Underwriter | Training-quality risk pricing for upstream operators |

---

## What's inside

13 CSV tables organized around `session_id` / `trainee_id` / `facility_id`
join keys: facility master → trainee master → training sessions → VR
movements (3D position + head rotation) → equipment interactions → alarm
events → emergency response → communication logs → evacuation sequences
→ safety violations → fatigue profiles → incident progression → pre-built
ML training labels.

| Table | Rows (sample) | What it represents |
|---|---:|---|
| `facility_master.csv` | 30 | 10-class facility × 10-region × VR environment version + operational complexity |
| `trainee_master.csv` | 500 | 8-class role × skill level × certifications + fatigue + stress susceptibility |
| `training_sessions.csv` | 2,500 | 20-class scenario × severity × fatigue × stress × completion score + grade |
| `vr_movements.csv` | 15,000 | 3D position (x, y, z) + head rotation (yaw, pitch) + hazard proximity + collision flag |
| `equipment_interactions.csv` | ~29,000 | 15 equipment types × 15 interaction types with correct-action flag + quality score |
| `emergency_response.csv` | ~14,800 | Multi-step response workflows with delay + success + containment status |
| `alarm_events.csv` | ~12,800 | ISA 18.2-aligned alarm priority + acknowledgment time + alarm flood flag |
| `communication_logs.csv` | ~16,200 | Communication type × clarity score × failure flag × command chain level |
| `evacuation_sequences.csv` | ~760 | Route × muster point × expected vs actual completion time |
| `safety_violations.csv` | ~810 | 10-class violation × severity × correction × coach intervention |
| `fatigue_profiles.csv` | ~7,500 | Per-session × 3-stage fatigue + reaction delay + cognitive load |
| `incident_progression.csv` | ~11,900 | Cascade staging × escalation probability × stabilization probability |
| `ai_training_labels.csv` | 2,500 | **Pre-built ML labels: 8 columns spanning hazard prob + response grade + containment + retraining flag + VR realism** |

Total: ~115,000 rows, ~14 MB. The full OIL-046 product is ~85 million
sessions and ~950 million VR movement records.

---

## Calibration sources

Every distribution and ratio is anchored to **named public references**.
Highlights:

- **IPIECA Competency Framework** — upstream operator competency
  classification and scenario taxonomy.
- **IOGP Process Safety Fundamentals** — facility classification and
  scenario severity bands.
- **IADC Well Control + WellCAP** — well-control training scenario
  taxonomy.
- **NFPA 1006** Technical Rescue Personnel Professional Qualifications —
  emergency responder training standards.
- **OPITO** Offshore Petroleum Industry Training Organization — VR-
  augmented offshore training requirements + session duration norms.
- **CCPS Process Safety + LOPA** — containment success benchmarks and
  rare-event drill scheduling.
- **NEBOSH** International General Certificate — safety violation
  taxonomy.
- **ISA 18.2 / EEMUA 191** — alarm priority bands and acknowledgment
  conventions.
- **UK HSE OHRA** + **API RP 755** — fatigue management applied to
  training environments.
- **DNV-RP-A203** simulator validation + emerging VR training-fidelity
  standards — realism scoring conventions.
- **ISO 14224:2016** — equipment classification compatible taxonomy.

---

## Validation scorecard

The wrapper ships a 10-metric scorecard (`validation_scorecard.json`) that
re-scores the dataset on every generation. Default seed 42 result:

| ID | Metric | Target | Observed | Source |
|---|---|---|---:|---|
| M01 | Facility-Type Taxonomy (floor) | ≥ 10 | **10** | IPIECA / IOGP |
| M02 | Scenario Taxonomy (floor) | ≥ 20 | **20** | IADC / IPIECA / NFPA 1006 |
| M03 | Equipment Taxonomy (floor) | ≥ 15 | **15** | ISO 14224 / IADC |
| M04 | Violation Taxonomy (floor) | ≥ 10 | **10** | NEBOSH / CCPS |
| M05 | Session Duration Median (min) | 30–90 | **65** | OPITO / IADC |
| M06 | Containment Success Rate | 0.65–0.85 | **0.727** | IPIECA / CCPS LOPA |
| M07 | Fatigue Exceedance Share | 0.06–0.18 | **0.110** | UK HSE OHRA / API RP 755 |
| M08 | VR Realism Score (mean, floor) | ≥ 0.87 | **0.918** | DNV-RP-A203 / OPITO |
| M09 | Rare-Event Label Rate | 0.005–0.045 | **0.022** | IPIECA / CCPS |
| M10 | Response Grade (mean) | 0.45–0.75 | **0.582** | IPIECA / IADC |

**Grade: A+ (100/100). Verified across seeds 42, 7, 123, 2024, 99, 1.**

Note: 6 of 10 metrics fall within 5% of target midpoint, and all 4 floor
metrics deliver complete taxonomy coverage at sample scale. The scorecard
is anchored to **11 distinct training-industry standards** spanning IADC,
IPIECA, OPITO, NFPA, CCPS, NEBOSH, ISA, UK HSE, API, DNV, and ISO — the
deepest standards-anchoring of any SKU in the catalog.

---

## Suggested use cases

- **Pass/fail prediction** — pre-built `training_pass_label` in
  `ai_training_labels.csv` enables binary classifier training for
  competency assessment.
- **Recommended retraining detection** — `recommended_retraining_flag`
  + `response_grade_score` × scenario_type supports retraining-recommender
  model training.
- **VR realism × performance correlation** — `vr_realism_score` per
  session × `procedural_accuracy` × `pass_label` enables simulator-
  fidelity ROI studies.
- **Multi-modal training event prediction** — join `alarm_events` +
  `communication_logs` + `equipment_interactions` + `vr_movements` on
  session_id to train multi-modal trainee-behavior models.
- **Fatigue-in-training analytics** — `fatigue_profiles` 3-stage scoring
  × session severity × procedural accuracy enables fatigue-aware training
  scheduling models.
- **Cascade-failure response training** — `incident_progression.csv`
  cascade staging × escalation probability × emergency response actions
  enables Bow-Tie / LOPA training-effectiveness modeling.
- **Equipment-interaction quality scoring** — per-interaction
  `correct_action_flag` × `actual_response_time_sec` vs `expected_response_time_sec`
  enables interaction-quality ML.
- **Evacuation timing prediction** — `evacuation_sequences.csv` expected
  vs actual completion time + route clear flag enables evacuation
  effectiveness modeling.
- **Cross-vertical immersive-training methodology** — the OIL-046
  generator architecture (20 scenarios × 15 equipment × VR movements ×
  labels) ports directly to Aviation, Maritime, Healthcare, Defense,
  Mining, and Manufacturing VR training research.

---

## Loading

```python
from datasets import load_dataset

trainees = load_dataset(
    "xpertsystems/oil046-sample",
    data_files="trainee_master.csv",
    split="train",
)
sessions = load_dataset(
    "xpertsystems/oil046-sample",
    data_files="training_sessions.csv",
    split="train",
)
labels = load_dataset(
    "xpertsystems/oil046-sample",
    data_files="ai_training_labels.csv",
    split="train",
)
```

Or with pandas directly:

```python
import pandas as pd
from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="xpertsystems/oil046-sample",
    filename="vr_movements.csv",
    repo_type="dataset",
)
df = pd.read_csv(path)
```

All 13 tables share these primary join keys:

- `trainee_id` → trainee_master ↔ sessions ↔ vr_movements ↔ equipment_interactions ↔ emergency_response ↔ communication_logs ↔ evacuations ↔ violations ↔ fatigue ↔ labels
- `facility_id` → facility_master ↔ trainee_master (home) ↔ sessions ↔ alarms ↔ evacuations ↔ labels
- `session_id` → sessions ↔ all event tables ↔ labels (1:1 or 1:N alignment)
- `incident_id` → emergency_response ↔ incident_progression (1:N cascade staging)

---

## Schema highlights

**`training_sessions.csv`** — `session_id`, `trainee_id`, `facility_id`,
`scenario_type` (20-class), `severity_level` ∈ {low, medium, high,
critical}, `severity_score`, `rare_event_flag`, `fatigue_score`,
`stress_score`, `mean_response_time_sec`, `procedural_accuracy`,
`communication_failure_flag`, `safety_violation_flag`,
`containment_success_flag`, `completion_score`, `training_grade`,
`ai_assist_enabled`, `vr_hardware_profile`.

**`vr_movements.csv`** — `movement_id`, `session_id`, `trainee_id`,
`timestamp`, `position_x`, `position_y`, `position_z`, `movement_vector`,
`head_rotation_yaw`, `head_rotation_pitch`, `proximity_to_hazard_m`,
`collision_or_trip_flag`, `safe_zone_flag`.

**`equipment_interactions.csv`** — `interaction_id`, `session_id`,
`trainee_id`, `equipment_id`, `equipment_type` (15-class),
`interaction_type` (15-class), `expected_response_time_sec`,
`actual_response_time_sec`, `correct_action_flag`,
`manual_override_flag`, `equipment_state_before/after`,
`interaction_quality_score`.

**`alarm_events.csv`** — `alarm_id`, `session_id`, `facility_id`,
`alarm_type`, `severity_level` (ISA 18.2), `acknowledged_flag`,
`acknowledgment_time_sec`, `alarm_flood_flag`, `false_alarm_flag`.

**`safety_violations.csv`** — `violation_type` (10-class:
wrong_valve_sequence, incorrect_ppe, missed_loto_step,
incomplete_permit_check, delayed_alarm_acknowledgment,
failed_communication_protocol, missed_gas_test, unauthorized_override,
+ 2 more), `procedure_breached`, `severity`, `coach_intervention_required_flag`,
`repeat_violation_flag`.

**`ai_training_labels.csv`** — pre-built ML labels:
`hazard_probability` ∈ [0, 1], `response_grade_score` ∈ [0, 1],
`operator_error_probability` ∈ [0, 1], `containment_success_label` (binary),
`emergency_escalation_label` (binary), `rare_event_label` (binary),
`training_pass_label` (binary), `vr_realism_score` ∈ [0, 1],
`recommended_retraining_flag` (binary).

---

## Calibration notes & limitations

In the spirit of honest synthetic data, a few things buyers of the sample
should know:

1. **Training pass rate is ~14% — much lower than industry-mature 60–80%.**
   The generator's `training_pass_label` requires containment success
   AND procedural accuracy AND low fatigue AND correct emergency response
   all combining; this multi-AND gate produces a low pass rate by design,
   biased toward identifying improvement opportunities. The scorecard
   validates the more useful **response_grade_score mean (0.58)** which
   sits in the IPIECA/IADC competency-development band. For pass-rate
   modeling work, threshold `response_grade_score > 0.65` directly to
   recover an industry-realistic ~60% pass rate.

2. **Operator error probability mean is ~71%.** Again, this is a training
   environment — operators are learning. Real-world (post-certification)
   operator error rates are much lower (~1–5%). For deployed-operator
   modeling, use OIL-038/039/040/045 which carry calibrated steady-state
   error rates.

3. **Recommended retraining flag ~86%.** This flag identifies *any*
   improvement opportunity, not just material competency gaps — most
   training sessions identify *something* to improve. For "actual
   retraining required" subset, intersect with `training_pass_label == 0`
   AND `response_grade_score < 0.50`.

4. **Violation severity is approximately uniform (~25% each across LOW /
   MEDIUM / HIGH / CRITICAL).** Industry-mature operations have pyramid-
   shaped violation distributions. The uniform distribution is intentional
   for balanced ML training; for pyramid-shaped sampling, use OIL-037
   (Regulatory Compliance) or OIL-045 (Workforce Safety Violations).

5. **VR movement data uses simple 3D position + head rotation.** No
   hand/controller pose data, no eye tracking, no biometric streams.
   For research requiring full VR biometric channels, the full product
   includes optional hand-tracking + eye-tracking + heart-rate streams.

6. **Equipment interactions assume binary correct/incorrect action.** Real
   operator training systems use graded correctness (e.g., "partially
   correct, sequenced wrong"). The full product carries a 5-tier
   correctness scale; sample uses the binary collapse.

7. **HF preview sizing** — default sample mode is 5K trainees × 25K
   sessions × 150K VR rows producing ~134 MB. The HF preview is reduced
   to 500/30/2,500/15,000, ~14 MB. All schemas, taxonomies, and scorecard
   calibrations are preserved at the smaller scale.

8. **Deterministic seeding.** All 13 tables are deterministic on `--seed`.
   Catalog default is seed 42. Seed sweep verifies Grade A+ across
   {42, 7, 123, 2024, 99, 1}.

---

## Commercial / full product

The full **OIL-046** product covers ~350,000 trainees × ~8,500 facilities
× ~85 million sessions × ~950 million VR movement records across a 5-year
horizon, with optional hand-tracking / eye-tracking / heart-rate biometric
streams, 5-tier graded correctness on equipment interactions, calibrated
industry-realistic pass-rate distributions, and configurable scenario-
portfolio composition for industry-specific competency stress testing.
Available under commercial license — contact
[pradeep@xpertsystems.ai](mailto:pradeep@xpertsystems.ai).

XpertSystems.ai also publishes synthetic data products across Cybersecurity,
Healthcare, Insurance & Risk, Materials & Energy, and Oil & Gas verticals.
Catalog: [huggingface.co/xpertsystems](https://huggingface.co/xpertsystems).