Datasets:

aiacontext
/

engquant

license: cc-by-4.0
language:
  - pt
pretty_name: EngQuant
size_categories:
  - n<1K
task_categories:
  - question-answering
  - text-generation
tags:
  - engineering
  - physical-quantities
  - quantitative-reasoning
  - structural-engineering
  - portuguese
  - benchmark
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/engquant.parquet

EngQuant

Adversarial benchmark of physical quantities in engineering, in Brazilian Portuguese.

EngQuant is a set of 800 procedurally generated, multi-step design-and-verification problems in engineering (Brazilian Portuguese), each with a verifiable numeric answer key per sub-quantity (gabarito). Every case is anchored in a primary bibliographic source — canonical textbooks, ABNT (Brazilian) technical norms, theses, and validated lecture notes — and stresses exactly where language models tend to fail on technical text: locale-specific numbers (decimal comma, thousands), compound dimensional units, normative/material identifiers, symbolic notation, regime decisions, and error propagation across steps.

Inclusion criterion: ≥ 4 auditable sub-quantities, ≥ 1 high-risk tag, a resolvable reference, and review_status = validated.

Usage

from datasets import load_dataset

ds = load_dataset("aiacontext/engquant", split="train")
print(ds[0]["title"])
# verifiable answer key per sub-quantity:
print(ds[0]["gabarito"])

Data fields

field	type	description
`id`	string	case identifier (e.g., `civ-balanco-cons-001`)
`title`	string	case title with the key quantities
`metadata`	struct	`discipline_folder`, `language`, `has_figure`, `has_table`, `n_subgrandezas`
`reference`	struct	bibliographic anchor (`source_id`, `chapter`, `section`, `page_range`, `notes`)
`tags`	list[string]	emergent stratification tags (cross-discipline)
`expected_signals`	list[string]	reasoning signals a correct solution should exhibit
`memoria_calculo`	struct	worked calculation memory
`expected_norms`	list[string]	normative references expected in the solution
`gabarito`	list[struct]	answer key: one auditable sub-quantity per entry (value + unit)
`prompts`	struct	prompt variants
`messages_for_api`	struct	chat-format messages

Tag stratification

Disciplines are organizational folders; statistical analysis stratifies by emergent tags that cut across disciplines. Counts over the 800 cases:

tag	n
`cadeia-erro-propagado`	800
`disciplina-civ-estrutural`	800
`identificador-simbolo-tecnico`	800
`locale-pt-br-decimal-virgula`	800
`identificador-material-br`	690
`identificador-norma-br`	690
`bibliografia-norma-abnt`	690
`locale-pt-br-thousands`	460
`constante-normativa`	420
`bibliografia-livro-br`	300
`decisao-regime`	200
`identificador-fragmentado`	120

Tag definitions are in tags_vocabulary.yaml; the bibliographic sources are in references.yaml; generation provenance (version, hashes, schema) is in manifest.json.

Provenance

Version 0.1.5, schema 1.0. Cases are generated procedurally (Latin-hypercube sampling over physically validated parameter ranges) and the answer keys are computed, not authored — each sub-quantity is numerically verifiable. The textual style of each case is informed by its bibliographic source; the cases themselves are original.

License

Citation

@misc{engquant2026,
  title        = {EngQuant: an adversarial benchmark of physical quantities in
                  Brazilian-Portuguese engineering},
  author       = {Leit\~ao Filho, Antonio de Sousa and Barros Filho, Allan Kardec Duailibe and
                  Lima, Fabr\'icio Saul and Santos, Selby Mykael Lima dos and
                  Sousa, Rejani Bandeira Vieira},
  year         = {2026},
  howpublished = {Hugging Face dataset},
  url          = {https://huggingface.co/datasets/aiacontext/engquant}
}