Datasets:

benchmarkanon
/

EngIntervene

pretty_name: EngIntervene
task_categories:
  - visual-question-answering
tags:
  - image
  - text
  - multimodal
  - benchmark
  - industrial-reasoning
configs:
  - config_name: default
    data_files:
      - split: test
        path: data/test-*.parquet

EngIntervene

Input release approved by the release maintainer. This repository contains all 3,229 items and 2,275 referenced input assets; scoring references are kept separately.

EngIntervene evaluates industrial visual understanding, relation modeling, diagnosis and reasoning about design changes. This anonymous-format copy contains all 3,229 items from the frozen benchmark; it is not a new collection and does not change the reference answers or scoring weights.

Tasks and coverage

Task	Capability	Items	Expected response
T1 Understand	Visual component and engineering-symbol understanding	2,149	Correct option ID(s)
T2 Model Relations	Entity, connection, path and mechanism reasoning	420	Answer and engineering relations
T3 Diagnose	Localizing and explaining issues and violated constraints	420	Diagnosis and supporting explanation
T4 Reason about Changes	Evaluating/proposing revisions and checking protected conditions	240	Revision, effect/mechanism and constraint reasoning

Stored domain ID	Domain	Items
`mechanical_industrial`	Mechanical Design, Manufacturing and Industrial Equipment	505
`robotics_mechatronics`	Robotics and Mechatronics	456
`automotive_integration`	Automotive System Integration	450
`aerospace_integration`	Aerospace Integration and Reliability	454
`building_mep`	Building Services and MEP Systems	460
`energy_equipment`	Power and Energy Equipment	454
`compact_electronics`	Compact Electronics Packaging	450

The single test split in this export means the complete 3,229-item evaluation corpus. It is not a training/adaptation split and is not just a historical challenge subset. No source-grouped train/validation/test experiment partition is invented by this export.

Loading locally

from datasets import load_dataset

dataset = load_dataset(
    "parquet",
    data_files={"test": "data/test-*.parquet"},
    split="test",
)
example = dataset[0]
images = example["images"]  # list of decoded PIL images
prompt = example["prompt"]

Run this snippet from the dataset directory, or use the included load_benchmark.py helper. You can also load the Hub version as follows. If repository visibility is private, first authenticate an account with access using hf auth login.

dataset = load_dataset("benchmarkanon/EngIntervene", split="test")

No credentials, private author identity or server paths are included in this repository.

Format

data/*.parquet: typed records with embedded image bytes. Images are a variable-length List(Image); their stored names are neutral basenames, never host filesystem paths.
metadata/questions.jsonl: matching text-only records and image IDs; image bytes remain in Parquet. This file alone is not the full multimodal input.
metadata/auxiliary_inputs.jsonl: original structured context / candidate_revisions fields, where present, joined by item_id. They are preserved as question-side auxiliary data; the historical evaluation prompt did not automatically append these fields.
metadata/assets.jsonl: image IDs, dimensions, pixel/file checksums, source links by ID and effective publication flags. Earlier source-metadata flags are retained in a separate historical field.
PUBLICATION_STATUS.json: anonymous maintainer publication decision and affected asset IDs.
metadata/sources.jsonl: third-party attribution and existing source-license metadata. Source institutions are not benchmark-author affiliations.
schema.json: the Hugging Face feature schema.
NOTICE.md: publication decision, source-attribution information and license-status distinctions; no blanket license is assigned by this update.

The typed columns include item_id, task, domain, subtask, system_family, question, choices, response_contract, prompt, image_ids, images, source_id, adaptation_type and redistribution_status.

The prompt field reproduces the existing task-level evaluation template, after any necessary identity redaction. It contains the benchmark question, options and response requirements; it does not include OCR archives, source-document excerpts, private labels, rubrics or backend-specific system prompts. Preserve all images and option IDs. T1 expects {"choice_ids":["<original option ID>"]}; T2–T4 expect {"answer":"..."}.

For comparison with existing runs, use prompt and all corresponding images. Adding auxiliary context, revision fields or other metadata to a model's input changes the evaluated input protocol and should be reported explicitly, rather than silently treated as the same baseline.

References and scoring

Gold references and rubrics are in a separate local scoring package, not in this input repository. A folder name such as private_scoring is not access control: never put the scoring package into a public input repository by mistake.

T1 uses exact option-ID/set matching. T2–T4 use the stored weighted rubric with applicable critical-error handling. T4 has two distinct metrics: partial-credit Atomic Rubric Score and Strict Revision Success. Preserve both. Changing an evaluator or a judge requires an explicit evaluation protocol; this packaging step runs no model and reports no new scores.

Anonymity and limitations

Private builder/reviewer identity fields, local account paths, build timestamps, infrastructure details, logs, credentials and repository history are excluded from the package. Image metadata is removed while preserving displayed pixels; images are not redrawn, cropped or anonymized by inventing content. Automated Chinese/English OCR and targeted inspection are used to screen disclosed identity strings. This does not prove that every unreadable, stylized or undisclosed identifier is absent.

Third-party source attribution remains. Do not erase source copyright notices to imply ownership. Source-language metadata is not a label for the normalized model-input language. Incomplete visual/relation taxonomy and unresolved semantic rubric issues are not silently relabeled by this release preparation.

Examples may contain other questions or source markings from original examination pages. The export preserves the current evaluated visual content. The benchmark mixes source-normalized, source-adapted and source-supported constructed questions; none of these terms means that every question is an untouched original examination question.

Format references

This export follows the documented Hugging Face image/Parquet representation and dataset-card structure. Format compatibility is not approval to redistribute source content.

Publication decision

The effective redistribution_status is approved_for_public_release_by_maintainer for every item. All referenced input assets have redistribution_allowed=true. This records the maintainer's explicit decision; it does not claim newly verified third-party permissions. Historical source flags and source-license metadata remain available. See PUBLICATION_STATUS.json and NOTICE.md.