mpvasilis/thema-panhellenic-exams

Thema

Thema is an open benchmark based on the Greek Panhellenic university entrance examinations (Πανελλαδικές Εξετάσεις ΓΕΛ). Models answer authentic exam questions in Greek. Responses are graded on the national 0 to 20 scale and can be converted to the admission points used by Greek university departments.

Website · Code · Leaderboard · Method

Dataset summary

	Current release
Exam years	2023 to 2026
Published subjects	10 of 10
Complete exam papers	40
Gradeable subquestions	519
Figure-dependent records	36
Total points per paper	100
Language	Greek
License	CC BY 4.0

The release is complete: all ten examined subjects across all four years.

Subject	Config	Field	2023	2024	2025	2026	Total	Points per section
Modern Greek Language and Literature	greek_lang	common	6	6	6	6	24	20/35/15/30
Ancient Greek	ancient_greek	1	11	11	12	13	47	10/40/50
History	history	1	6	6	6	6	24	25/25/25/25
Latin	latin	1	11	10	10	10	41	20/20/30/30
Mathematics	math	2	16	16	16	14	62	25/25/25/25
Physics	physics	2	17	17	16	16	66	25/25/25/25
Chemistry	chemistry	2	16	15	16	16	63	25/25/25/25
Biology	biology	3	18	18	17	18	71	25/25/25/25
Informatics	informatics	4	17	18	17	17	69	25/25/25/25
Economics	economics	4	13	14	12	13	52	25/25/25/25
All			131	131	128	129	519

Question types across the 519 records: open 237, mcq 90, numeric 76, code 50, true_false 34, essay 32.

Point distribution and its exceptions

Every paper sums to exactly 100 points, but the widely assumed "25 points per section" rule holds only for the science, informatics, economics, and history papers. Three subjects have their own fixed structure, stable across all four years, and any validator must special-case them:

• Modern Greek Language and Literature (greek_lang): 20 / 35 / 15 / 30. Section B carries three subquestions (15 + 10 + 10) and section D is a single 30-point composition.
• Latin (latin): 20 / 20 / 30 / 30. Translation 20, introduction and etymology 20, morphology 30, syntax 30.
• Ancient Greek (ancient_greek): 10 / 40 / 50, three sections rather than four. Comprehension of the set text 10, interpretation and parallel text and etymology and literary history 40, unseen passage 50, of which the translation alone is 20.

The invariant that holds everywhere is sum(max_points) == 100 per (subject, year).

Why this benchmark

Real examinations with real consequences

The Panhellenic exams determine admission to Greek public universities. A score can therefore be expressed in a familiar national scale and mapped to university departments, subject to the correct department weights, the Minimum Admission Threshold, and the cutoff for a given year.

Native Greek evaluation

The prompts, worked solutions, and essays are evaluated in Greek. The benchmark covers mathematical, scientific, computational, economic, historical, and language tasks without translating them into English. Four of the ten subjects concern the Greek language or its classical ancestors, and several task types have no English equivalent at all: a 20-point translation of an unseen Ancient Greek passage, active-to-passive syntactic conversion in Latin answered in Greek, and source-based history essays graded against specific chapters of a specific state textbook.

Partial credit

Each record includes a step-level scoring rubric. Multi-step answers can receive partial credit rather than being reduced to exact-match accuracy.

Annual refresh

A new split is added after each June examination cycle. Older years remain useful for development and retrospective analysis. The newest year is the preferred test split for models whose training cutoff predates publication of that year's papers.

Controlled reasoning-effort study

Thema includes a pre-registered study that holds the base model fixed and changes only the reasoning-effort setting:

Base model	Conditions	Runs per condition	Primary outcome
GPT-5.6 Sol	none, low, medium, high, max	3	Exam grade and admission-point change

Secondary outcomes are latency, output tokens, run-to-run variation, and score gained per additional unit of compute. Results are published only after all three complete runs for a condition have been graded. The five condition files are not yet complete, so no effort curve is reported in this card.

Current comparable results

The current 2026 release compares 13 models on Modern Greek, Mathematics, and Physics. Every listed score is the mean of three runs on the same question set. Modern Greek and Mathematics cover all 6 and 14 questions respectively; Physics covers 13 of 16 questions, with the three figure-dependent questions excluded uniformly. No admission-point total is reported yet because a complete four-subject field is not currently available.

Model	Modern Greek	Mathematics	Physics	Mean grade
Claude Fable 5	19.77	19.33	20.00	19.70
GPT-5.6 Sol	19.23	19.77	20.00	19.67
Gemini 3.1 Pro	18.87	19.87	20.00	19.58
Grok 4.5 (high)	18.57	19.42	19.91	19.30
Claude Sonnet 5	18.10	19.67	20.00	19.26
GLM-5.2 (high)	18.93	18.62	19.91	19.15
GPT-5.2	18.43	18.67	20.00	19.03
DeepSeek V4 Flash	18.47	18.63	19.83	18.98
Claude Haiku 4.5	17.40	18.27	20.00	18.56
DeepSeek V4 Pro	18.43	17.47	19.57	18.49
Gemma 4 31B	18.13	16.77	20.00	18.30
GPT OSS 120B	16.70	15.33	20.00	17.34
MiniMax M3	13.30	15.27	19.05	15.87

Beyond this comparable table, 80 completed model–paper result sets are available in the website evidence feed. Partial grading now exists in Ancient Greek, Biology, Chemistry, Economics, and Informatics, but those papers remain outside the ranking until all non-figure questions have three settled runs on a common set.

Load the data

from datasets import load_dataset

math_2026 = load_dataset(
    "mpvasilis/thema-panhellenic-exams",
    "math",
    split="2026",
)

print(math_2026[0])

Select another subject through the configuration name and another exam year through the split name.

Solutions and scoring guides

Solutions are included directly in every dataset record. All 519 records contain:

• official_answer: Thema's complete reference solution in Greek;
• rubric: the step-level marking guide, including partial-credit allocation;
• max_points: the question's contribution to the 100-point exam paper.

from datasets import load_dataset

row = load_dataset(
    "mpvasilis/thema-panhellenic-exams",
    "math",
    split="2026",
)[0]

print(row["prompt"])
print(row["official_answer"])
print(row["rubric"])

These are original Thema reference answers and rubrics. Third-party publisher answer-key PDFs used for factual verification are not redistributed.

Record schema

One JSON object represents one gradeable subquestion.

Field	Type	Description
year	integer	Examination year
subject	string	Stable subject identifier
field	integer	Associated academic field
thema	string	Exam section, usually A, B, C, or D
question_id	string	Globally unique question identifier
prompt	string	Full Greek prompt, with LaTeX where needed
max_points	integer	Maximum points for the subquestion
official_answer	string	Reference answer written for this release
rubric	string	Step-level scoring guide and partial-credit rules
question_type	string	mcq, true_false, numeric, open, essay, or code
requires_image	boolean	Whether the original question depends on a figure
image_path	string or null	Relative figure path when available

For every (subject, year) pair, max_points sums to 100. Every question_id is unique.

How the Greek admission system works

Candidates sit four written subjects tied to an orientation group and academic field. Each paper is marked out of 100 and converted to the 0 to 20 national scale. University departments assign weights to the four subjects, with weights summing to 100 percent.

For a specific department:

admission points = 1,000 × Σ(subject grade × department weight)

The maximum is 20,000 points. Eligibility also depends on the Minimum Admission Threshold. Final department cutoffs change every year because they reflect candidate preferences and the number of available places.

Department-level weights and cutoffs must be matched to the correct year. The provisional coefficient table in the project repository requires official verification before school-admission mappings are treated as final research results.

Evaluation protocol

1. Generation. Each model answers the original Greek question without retrieval or external tools unless a track explicitly states otherwise. Temperature is 0.
2. Repeated runs. The default protocol uses three independent runs per question.
3. Closed-form grading. Multiple-choice, true-or-false, and numeric responses are checked programmatically.
4. Worked-response grading. Open responses are scored against the question rubric. A human-graded sample is used to calibrate and audit this stage.
5. Subject score. Awarded points sum to a score out of 100, then convert to the national 0 to 20 scale.
6. Admission mapping. Four subject grades are combined using department-level weights. Eligibility and annual cutoffs are applied separately.

The website exposes completed run scores, response records, scoring rationales, latency, token counts, and error labels where available.

Data sources and validation

The original exam papers are public documents released by the Greek Ministry of Education. Thema does not redistribute the source PDFs through this dataset. Prompts are transcribed into structured records, and every prompt is self-contained: set texts, unseen passages with their glossaries, parallel texts, Latin texts, and historical source extracts with their citations are all embedded in the record.

Reference answers and rubrics are written for Thema. Published indicative solutions are used for factual cross-checking where available, but copyrighted solution prose is not copied into the release. Cross-checking covers facts only: correct option letters, true/false characterisations, final numeric values, grammatical and syntactic forms, dates, names, and quantities.

Verification sources by subject:

• Nine subjects were verified against the ΟΕΦΕ indicative answers published on the day of each examination.
• History has no publicly retrievable ΟΕΦΕ key. It was verified against the ΔΙΑΚΡΟΤΗΜΑ indicative answers and, as the primary authority, grounded point by point in the official state textbook «Θέματα Νεοελληνικής Ιστορίας» (Institute of Educational Policy / ΙΤΥΕ Διόφαντος, code 22-0163-02, ebooks.edu.gr). All 24 history records carry an explicit chapter-and-section citation block in official_answer, and each rubric distributes points across named content points rather than dividing them evenly. The source-based questions additionally carry a block stating which content points come from the textbook and which come only from the given source extracts. The textbook is cited, not reproduced.
• Latin 2023 and 2024 have no external key and rest on our own philological analysis; the specific items worth human confirmation are named in the repository validation log.

All 40 files pass structural checks:

• valid JSON on every line;
• exactly 100 available points per paper;
• unique question identifiers across the whole dataset;
• valid question types;
• consistent requires_image and image_path values.

The per-subject, per-year validation log with confidence ratings, the corrections the verification pass produced, and the documented divergences from external keys is data/processed/VALIDATION.md in the project repository. Open-ended humanities answers and figure-dependent questions still need additional human review before final leaderboard publication.

Contamination policy

Public exam papers can enter training corpora after release. Results should therefore be interpreted against each model's training cutoff.

• Use 2023 to 2025 for development and retrospective comparisons.
• Prefer 2026 as the current held-out split for models with a training cutoff before the June 2026 exam publications.
• Report the exact model identifier, training cutoff when known, backend, prompt, effort setting, and run count.
• Do not combine contaminated and post-cutoff splits into a single headline score without separate reporting.

Limitations

• Open-response grading. Essays and worked solutions have more grading uncertainty than closed-form questions.
• Figures. Thirty-six records require diagrams that are not yet packaged with the public dataset. They are flagged requires_image: true and are excluded from scoring rather than answered without the figure. The affected subjects are Physics (16), Biology (10), Chemistry (7), and Informatics (3); the other six subjects have none.
• Coverage. All ten subjects and all four years are published. Coverage of the evaluation side is narrower than coverage of the dataset: not every subject has been sat by every model.
• Admission mapping. Department weights, the Minimum Admission Threshold, and cutoffs are year-specific. Provisional mappings are clearly labeled on the website.
• Current results. The comparable leaderboard covers the June 2026 Mathematics and Physics papers across seven models, with three runs per question. Twenty-one complete model–paper result sets are available in the evidence feed, while partial results in five additional subjects remain outside the ranking. No full scientific field is currently complete, so admission points are intentionally blank. Judge grading is blind to the answering model, but the judge and some graded models are from the same family, and no human calibration sample has been scored yet.

Responsible use

This dataset is intended for research, evaluation, and educational analysis. Scores do not establish that a system can replace a student, teacher, examiner, or admissions process. Model output may reproduce errors confidently and should not be used as study guidance without expert review.

Citation

@misc{thema2026,
  title  = {Thema: The Greek National Exam Benchmark},
  author = {Balafas, Vasilis and Thema contributors},
  year   = {2026},
  url    = {https://huggingface.co/datasets/mpvasilis/thema-panhellenic-exams}
}

Contact and contributions

Use the project repository for corrections, evaluation submissions, and reproducibility issues. Contributions should identify the source paper, exam year, subject, and affected question_id.