SteelBench: A Diagnostic Benchmark for Vision-Language Models in Industrial Safety Monitoring

SteelBench is a diagnostic benchmark of densely annotated CCTV clips from an operating integrated steel plant. It is designed to evaluate vision-language models (VLMs) on real-world industrial action recognition, PPE assessment, and safety-violation detection — under naturally occurring degradation (dust, glare, steam, low light), at distances and crowdedness levels that curated benchmarks miss.

The benchmark is paired with an audit protocol that empirically bounds the influence of model-assisted annotation on evaluation results. The protocol is itself a methodological contribution: a reusable recipe for provenance-aware dataset construction in domains where annotation is scarce or expensive.

At a glance

Clips	1,345 (15 s each, 1080p, H.264)
Frames	10,760 evenly-spaced JPEGs (8 per clip) at quality 95
Sites	16 across an integrated steel plant (ASP, BF, CRM, RED, RERS, SMS, etc.)
Action taxonomy	25 codes in 6 groups (A–F) + X1 (unlisted)
PPE items	5 per worker (helmet, high-vis vest, safety shoes, welding protection, harness)
Safety rules	55 general (UA-G) + site-specific (UA-SP, UA-RED, UA-CRM, …)
Visibility conditions	6 (clear, steam, dust, smoke, low_light, glare)
Annotation layers	Layer 1 (scene-level, >5 workers) / Layer 2 (per-person, ≤5 workers)
Annotators	5 active tier-1 annotators + 2 experts + 1 safety officer
Ground truth provenance	Highest-priority annotation per clip (expert > tier-1 > safety_officer)
License	CC-BY-NC 4.0 (non-commercial, attribution)

Dataset structure

SteelBench/
├── README.md                 # this file
├── LICENSE                   # CC-BY-NC 4.0 full text
├── ethics.md                 # surveillance consent + face anonymization rationale
├── croissant.json            # Croissant Core + RAI metadata
├── data/
│   ├── frames/               # 1,345 dirs × 8 .jpg = 10,760 jpg (~11 GB, anonymized)
│   ├── annotations/          # 1,345 canonical GT JSONs
│   ├── annotation_source.json  # per-clip provenance map
│   └── safety_review/        # 186 safety officer reviews (parallel layer)
├── manifests/
│   ├── gt_clips.json         # canonical 1,345 clip_id list
│   ├── batch_manifest.csv    # per-clip metadata (site, work_area, BRISQUE, etc.)
│   ├── camera_zones.csv      # zone tag per camera_id
│   └── safety_rules.yaml     # rule definitions
├── eval_data/
│   ├── prompt_sensitivity_clips.json   # 150-clip ablation subset
│   └── ablation_150_clips.json
└── sample/                   # 50-clip stratified preview (594 MB) for reviewers
    ├── clips/                # 50 .mp4 (full original clips for sanity-check)
    ├── frames/                # 400 anonymized .jpg
    ├── annotations/           # 47 canonical GT JSONs (3 of 50 lacked annotations
    │                          #  in the canonical set; documented for transparency)
    └── sample_manifest.csv

The sample/ directory satisfies the NeurIPS Datasets & Benchmarks track requirement that >4 GB datasets ship a small sample for reviewer inspection.

Why no full mp4 clips in the main release? The 8 representative jpgs per clip were anonymized via face blurring; the full mp4s were not (re-encoding 360+ frames per clip with face detection was out-of-scope for this release and the camera-distance argument that justifies the low face-detection rate on jpgs becomes less reliable across continuous video where movement reveals more). The 50-clip sample/ subdir does include mp4s for reviewer inspection — this is a small, scoped exposure consistent with the double-blind review process. For full mp4 access for legitimate research, contact the authors after acceptance.

How to load

from huggingface_hub import snapshot_download
import json

# Full dataset
local_dir = snapshot_download(repo_id="steelbench/SteelBench", repo_type="dataset")

# Just the manifest + annotations (no media)
local_dir = snapshot_download(
    repo_id="steelbench/SteelBench",
    repo_type="dataset",
    allow_patterns=["manifests/*", "data/annotations/*", "README.md", "LICENSE"],
)

clip_ids = json.load(open(f"{local_dir}/manifests/gt_clips.json"))
ann = json.load(open(f"{local_dir}/data/annotations/{clip_ids[0]}.json"))

Datasheet for Datasets

Motivation

For what purpose was the dataset created? SteelBench was created to fill a gap in VLM evaluation: existing video-and-action benchmarks (Kinetics, ActivityNet, Charades) are curated, well-lit, and unambiguous; existing industrial datasets (IndustryEQA, MonitorVLM, iSafetyBench) are simulated, synthetic, or single-task. We needed a real-deployment benchmark with multiple evaluation dimensions (perception, safety reasoning, calibration) to test whether modern VLMs are deployment-ready in industrial monitoring.

Who created the dataset? The dataset was created by the SteelBench authors as part of an academic research project. (Author identities withheld during double-blind review.)

Funding / interests: No commercial relationship to the steel plant. Footage shared under research-only data-use agreement.

Composition

What does each instance represent? A 15-second clip from a fixed-position CCTV camera in a steel plant operational area. Each clip is annotated with: scene-level action labels, per-person action codes (when ≤5 workers), PPE assessment per worker, safety rule citations (when violations are observed), spatial context tags, visibility conditions, and an annotator-confidence score.

How many instances are there? 1,345 clips total. Per-site distribution ranges from 1 (TAR Plant) to 211 (CRM 1&2).

Does the dataset contain all possible instances or is it a sample? Sample. Source video totals ~149 hours from 117 unique videos; SteelBench is a curated subset stratified for action-class balance and visibility diversity. Curation pipeline is open-sourced in the companion code repository (extract_clips.py, filter_clips.py, curate_batch.py).

What data does each instance consist of?

• A 15-second .mp4 clip (1080p, H.264)
• 8 evenly-spaced JPEG frames (anonymized — see ethics.md)
• A canonical GT annotation JSON with the structured fields documented in annotation_tool/schema_validator.py

Is there a label or target? Yes — per-clip structured annotation with multiple targets: action codes, PPE compliance, safety violations, scene type, worker count, visibility conditions.

Are there labeled subsets / splits?

• manifests/gt_clips.json — full 1,345-clip benchmark
• eval_data/prompt_sensitivity_clips.json — 150-clip stratified ablation subset (used in Section 7 prompt-sensitivity ablation in the paper)
• eval_data/ablation_150_clips.json — 150-clip stratified subset for frame-density ablation
• sample/ — 50-clip preview for review

Are there missing modalities or relationships between instances? Different clips may share a source video and camera_id; this is documented in manifests/batch_manifest.csv.

Collection process

How was the data acquired? CCTV footage from an operating integrated steel plant, streamed continuously from 64 fixed cameras across 16 work areas. Source videos cover December 2025–April 2026.

What sampling/processing was applied?

• Person detection (YOLOv8-n) on 0.5 fps sampled frames
• Detection-interval merging with 5 s gap tolerance and 2 s padding
• 15 s fixed-window slicing
• Quality filtering (BRISQUE, person-detection ratio, bounding-box area)
• Stratified curation by action class and visibility condition

Who was involved in data collection? Plant safety personnel installed and maintain the camera infrastructure. The research team applied processing and curation. The annotation pipeline involved 5 trained tier-1 annotators, 2 domain experts (industrial safety), and 1 safety officer.

Over what time frame was the data collected? Source footage: 2025-11–2026-04. Annotation: 2026-04 onwards.

Preprocessing / cleaning / labeling

Was the data preprocessed/cleaned/labeled? Yes. See the curation pipeline above and the annotation tooling in annotation_tool/ (schema_validator.py, safety_rules.py, app.py).

Was the "raw" data saved in addition to the preprocessed/cleaned/labeled data? Raw source videos are retained by the data provider but not released; this dataset ships the curated 15 s clips only.

Annotation methodology — model-assisted with audit: The annotation tool pre-fills the structured form using a single VLM (Qwen3-VL-30B-A3B) so annotators verify rather than write from scratch. The paper's audit protocol (Section 4) empirically bounds the influence of this pre-fill via override rate, direction analysis, and dual-track calibration (anchored vs blind). The full audit code and intermediate audit data are in the companion code repository.

Uses

For what purposes can the dataset be used?

• Evaluation of VLMs on industrial action recognition, PPE detection, and safety-violation reasoning
• Benchmarking calibration and degradation robustness
• Research on annotation provenance and audit protocols

What restrictions apply?

• Non-commercial use only (CC-BY-NC 4.0)
• No re-identification attempts
• No use for surveillance product training without separate agreement
• See ethics.md for full ethics statement

Are there tasks for which the dataset should NOT be used?

• Worker identification or biometric profiling (faces are blurred)
• Predictive surveillance of protected categories
• Production deployment without further validation (the paper shows no evaluated VLM is deployment-ready: best DRS = 0.40)

Distribution

Will the dataset be distributed? Yes, on Hugging Face under CC-BY-NC 4.0.

Is there an associated paper? Yes, NeurIPS 2026 D&B submission. (Cite once accepted; pre-acceptance citation withheld during review.)

Maintenance

Who is supporting/hosting/maintaining the dataset? The author team. Issues and questions should be filed on the companion code repository.

Are there errata? Will be tracked as GitHub issues in the code repo post-acceptance.

Will old versions be supported/hosted? Yes — Hugging Face dataset versions are tagged.

Known limitations

• Single facility: All clips are from one integrated steel plant. The paper claims internal generalization across 16 work areas but does not claim cross-facility generalization.
• Annotation provenance: 25.6% of GT comes from domain experts; 74.4% from tier-1 annotators. The annotation source for each clip is exposed in data/annotation_source.json.
• Scene-level vs per-person: Clips with >5 workers receive only scene-level (Layer 1) annotations; per-person assessment (Layer 2) applies to ≤5-worker scenes.
• Single annotation per clip: Most clips have one annotator. 105 clips are doubly annotated by both experts (used for inter-expert agreement computation in the paper).
• Anonymization is best-effort: MediaPipe long-range face detection identified faces in 4.81% of frames and blurred them. The remaining 95.19% had no detectable face — but small / partially-occluded faces may have been missed. See ethics.md.
• Action-class long tail: Some classes (A2, A3, B4, C2, C4, D3, X1) have <30 instances. Per-class metrics in the paper acknowledge this.

Citation

@inproceedings{steelbench2026,
  title     = {SteelBench: Evaluating Vision-Language Models in Real-World
               Industrial Environments},
  author    = {Anonymous Authors},
  booktitle = {NeurIPS Datasets and Benchmarks Track},
  year      = {2026},
}

License

This dataset is released under Creative Commons Attribution-NonCommercial 4.0 International (CC-BY-NC 4.0). Code in the companion repository is released under Apache-2.0.