README.md

metadata

license: mit
language:
  - en
pretty_name: JEE Advanced 2022–2026
size_categories:
  - n<1K
task_categories:
  - question-answering
  - visual-question-answering
tags:
  - benchmark
  - exam
  - jee
  - jee-advanced
  - mathematics
  - physics
  - chemistry
  - multimodal
  - latex
configs:
  - config_name: default
    data_files:
      - split: test
        path: data/test-*.parquet

JEE Advanced 2022–2026

All 510 questions from the ten JEE Advanced papers of 2022–2026 (Paper 1 and Paper 2 of each year), in English. Each question comes with:

• its official answer from the final answer key
• its marking scheme
• the figures it needs: 279 images, used by 155 questions

JEE Advanced is the entrance exam of the Indian Institutes of Technology (IITs). Its Mathematics, Physics and Chemistry questions need multi-step reasoning, and many depend on a diagram, a graph or a chemical structure. That makes the dataset a hard benchmark for text and vision-language models.

Text is Markdown with LaTeX ($...$ and $$...$$). Every string renders with KaTeX.

Contents

year	paper	Mathematics	Physics	Chemistry	questions	max marks
2022	1	18	18	18	54	180
2022	2	18	18	18	54	180
2023	1	17	17	17	51	180
2023	2	17	17	17	51	180
2024	1	17	17	17	51	180
2024	2	17	17	17	51	180
2025	1	16	16	16	48	180
2025	2	16	16	16	48	180
2026	1	16	16	16	48	180
2026	2	18	18	18	54	180

type	questions
numerical	129
multi_correct	120
single_correct	108
integer	72
matrix_match	57
single_digit_integer	24

What each type means:

• single_correct: exactly one of options A–D.
• multi_correct: one or more of A–D, with partial marks.
• matrix_match: match List-I to List-II, then pick one of options A–D.
• integer and single_digit_integer: a non-negative integer.
• numerical: a real number, usually to two decimal places.

36 questions are paragraph-based: groups of questions share one passage.

Quick start

import json
from datasets import load_dataset

ds = load_dataset("propar2750/jee-advanced", split="test")
q = ds[0]
print(q["question_text"])
answer = json.loads(q["correct_answer"])   # e.g. {"kind": "option", "value": ["A", "C"]}
for name, img in zip(q["image_ids"], q["images"]):   # PIL images
    print(name, img.size)                            # name matches the [FIG_1]-style placeholder in the text

Fields

field	type	description
id	string	{year}_P{paper}_{subject}_Q{nn}, e.g. 2024_P2_Physics_Q07
year, paper	int	2022–2026; 1 or 2
subject	string	Mathematics, Physics or Chemistry
q_number, section	int	question number within the subject, and the paper section it is in
type	string	see the table above
is_paragraph, paragraph	bool, struct	for paragraph-based questions, the shared passage (id, label, text); otherwise false and null
question_text	string	the question (Markdown + LaTeX), with figure placeholders such as [FIG_1]
images	list of images	the figures, as PNG crops of the paper
image_ids	list of strings	the name of each image, in the same order. FIG_1… are question figures, PFIG_1… passage figures, FIG_A… option figures, and FIG_P… list-entry figures. Each name appears as [name] in the text, options or passage at the point where it is printed
options	struct	choices (label A–D and text); for matrix_match, also list_I and list_II. Null for numeric types
marking	struct	marks for a full answer, zero (unanswered) and negative (wrong). partial holds the paper's partial-marking rules (marks, condition)
answer_kind	string	option, number, range, any_of or bonus; the same as the kind inside correct_answer
correct_answer	JSON string	the official answer, see below
answer_raw	string	the answer exactly as the official key prints it
source	struct	the PDF page(s) the question is on, the extraction model, and the prompt version
review	struct	how the question was checked: status, reviewer, flags and notes, see Review status

Answers

correct_answer is stored as a JSON string because its shape depends on its kind:

kind	example	correct when
option	{"kind": "option", "value": ["A", "C"]}	the chosen set of options equals value
number	{"kind": "number", "value": 3}	integer types: exactly value. numerical: within ±0.01
range	{"kind": "range", "min": 2.35, "max": 2.37}	min ≤ answer ≤ max, inclusive
any_of	{"kind": "any_of", "value": [<answer>, ...]}	the answer matches any one of the alternatives. Used for 16 questions where the key accepts more than one answer
bonus	{"kind": "bonus"}	the question was dropped ("marks to all"), see below

7 questions were dropped by the official key, which gives marks to every candidate: 2022_P1_Physics_Q15, 2022_P2_Physics_Q14, 2022_P2_Physics_Q17, 2023_P2_Physics_Q16, 2023_P2_Physics_Q17, 2024_P2_Chemistry_Q11, 2025_P2_Physics_Q04.

• Measuring accuracy: leave them out, e.g. ds.filter(lambda q: q["answer_kind"] != "bonus").
• Reproducing official paper totals: award them full marks.

For multi_correct, partial marks follow the paper's own rules, stored in marking.partial as each rule's marks and condition.

How it was built

1. The official question papers and final answer keys were downloaded as PDFs.
2. Each page was rendered and transcribed by qwen/qwen3-vl-235b-a22b-instruct, using prompt v6.
3. z-ai/glm-4.6v transcribed every page independently as a second opinion.
4. Figures were cropped from the PDF, using the PDF's embedded image boxes where available.
5. The answer keys were parsed. For 2023–2026 they were read from the coloured answer text in the PDFs. For 2022, they were read from page images by a vision model, because that key is partly scanned.
6. Automatic checks then compare the two transcriptions with each other and with the PDF's text layer. They also verify answer format, option count, figure placeholders, marking against the printed section instructions, and that every string renders in KaTeX.

Any question that failed a check was reviewed and corrected against the page images. Corrections are applied on top of the extraction, so they survive a rebuild.

The extraction models are open-source models only.

Review status

review.status says who checked each question:

status	questions	meaning
auto_ok	349	passed every automatic check, and both models agreed. Not individually reviewed
model_verified	160	failed a check, then reviewed and corrected where needed by a model (review.reviewer, e.g. claude-opus-5-5) against the page images and the key
human_verified	1	checked by a person

Most questions have not been checked by a human. For flagged questions, review.flags lists the checks that failed before review; review.notes lists informational messages. If you find an error, please open a discussion on this repository.

Raw files

jsonl/ holds the database exactly as built:

• questions.jsonl: one record per question. correct_answer is a JSON object, and images are listed as {id, path}.
• images/: the image files.
• manifest.json: counts, models and the build date.

Use jsonl/ if you prefer plain files over Parquet.

Limitations

• Transcription. The text and LaTeX are model transcriptions of the printed papers. Most questions were never compared with the page by a human, so typos are possible.
• Figures. Figures are crops of the paper. Chemical structures and graphs are images, not text.
• The 2022 answer key was read from scanned pages by a vision model.
• Contamination. The papers and their solutions have been public since each exam, so models trained on web data may have seen them. Newer papers are less likely to have been seen.

License and source

The questions, answer keys and figures come from the official JEE Advanced 2022–2026 papers and final answer keys, published by the IITs that organise the exam. They remain the work of those institutes.

The transcriptions, structure and annotations in this dataset are released under the MIT license.

Built 2026-10-26.

Citation

@misc{jee_advanced_2022_2026,
  title        = {JEE Advanced 2022--2026: questions, official answers and figures},
  author       = {propar2750},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/propar2750/jee-advanced}}
}