Datasets:

Axilo67
/

openstem-qa

pretty_name: OpenSTEM-QA
tags:
  - stem
  - mathematics
  - physics
  - chemistry
  - synthetic
  - multiple-choice
  - jee-inspired
task_categories:
  - question-answering
language:
  - en

OpenSTEM-QA

10,000 original, synthetic STEM multiple-choice QA records by Axilo Labz.

Hugging Face target repository: Axilo67/openstem-qa

Release contents

10,000 unique question records.
Exactly four options (A–D) per question.
One designated answer option.
A worked numerical solution.
An explanation field for each option.
3,340 mathematics, 3,330 physics and 3,330 chemistry questions.
train (9,000), validation (500) and test (500) JSONL splits.
A combined JSONL file containing all records.

Subject and topic coverage

Mathematics

Quadratic-root identities, arithmetic and geometric progressions, polynomial differentiation, polynomial definite integrals, combinations, probability without replacement, determinants, vector dot products, and complex-number division.

Physics

Projectile motion, work–energy, friction on an inclined plane, rotational dynamics, series-parallel circuits, Coulomb's law, capacitor energy, isochoric ideal-gas processes, thin-lens magnification, and mass–spring simple harmonic motion.

Chemistry

Limiting reagents, ideal gases, strong-acid/strong-base mixtures, first-order kinetics, electrolysis, colligative properties, chemical equilibrium, the Nernst equation, degree of unsaturation, and formal oxidation-state charge balance in hypothetical coordination ions.

Generation method

This release was generated with deterministic Python templates and randomized numerical parameters. It was not generated by zai-org/GLM-5.3-Flash. The hosted GLM model was not run because the available Hugging Face remote-compute route returned 402 Payment Required.

The questions are newly worded and parameterized. They are inspired by the subject matter and calculation style of Indian JEE preparation, but they are not official JEE questions and do not reproduce past-paper questions.

Record schema

Each JSONL object contains:

Field	Meaning
`id`	Unique record identifier
`subject`, `topic`, `subtopic`	Curriculum labels
`difficulty`	Provisional `hard` or `very_hard` label
`difficulty_labeling_method`, `difficulty_note`	Explain the provisional label
`question_type`	`single_correct_mcq`
`question`	Question statement
`options`	Exactly four choices keyed A–D
`correct_answer`	Correct option label
`correct_answer_value`	Numeric answer with units where applicable
`solution`	Worked calculation
`option_explanations`	Explanation for all four choices
`concepts_required`	Concept tags
`reasoning_steps`	Template-declared step count
`answer_type`, `language`	Answer format and language
`origin`, `source_type`	Synthetic-data provenance
`generator`, `generator_template_id`	Generator and question family
`verification_status`, `verification_note`, `answer_validation`	Validation provenance and limitations

Validation performed

The generated release passed these structural checks:

Exactly 10,000 rows.
10,000 unique IDs.
10,000 unique question strings after basic punctuation/case normalization.
Exactly four option keys per record (A, B, C, D).
Four distinct option strings per record.
The correct_answer label points to the option that matches correct_answer_value.
A non-empty question and worked solution.
An explanation field for every option.
Exact split sizes: 9,000 / 500 / 500.
Balanced subject representation across splits.

The validation report is included as quality_report.json.

Important limitations

Not expert-reviewed: Structural checks confirm schema and answer-key consistency. They do not replace an independent subject-matter solution review.
Difficulty is provisional: This release labels records hard or very_hard by question-template family. It does not claim examiner-calibrated JEE Advanced difficulty. No records are labeled ultra_hard in v0.1 because the current template set does not justify that claim.
Parameterized families: The 10,000 rows come from 30 underlying templates. Values and wording vary, but solution structures repeat. This is not equivalent to 10,000 independently authored problems.
Evaluation leakage: The same template families appear in train, validation and test. The validation/test splits are useful for basic pipeline checks, but should not be treated as template-generalization benchmarks.
Numeric emphasis: Most questions are calculation-based. Proof-based mathematics, multi-select questions, assertion-reason questions, qualitative chemistry, organic reaction mechanisms, and experimental-data interpretation are limited or absent.
Distractors: Incorrect numerical options are distinct alternatives. Their explanations generally identify the computed value rather than diagnosing a specific misconception.
Licensing: No separate reuse license is asserted in this seed release. A formal license should be selected before third-party redistribution or commercial use.

Loading

from datasets import load_dataset

ds = load_dataset(
    "json",
    data_files={
        "train": "data/train.jsonl",
        "validation": "data/validation.jsonl",
        "test": "data/test.jsonl",
    },
)
print(ds)

To use a single combined file:

from datasets import load_dataset
ds = load_dataset("json", data_files="openstem_qa_v0.1.jsonl", split="train")

Intended use

Useful for experiments in STEM question answering, numeric MCQ supervision, answer extraction and worked-solution training. Independently review examples before using them in high-stakes educational settings.

Axilo Labz — Exploring Intelligence. Engineering Possibilities.