medresetuzzehra/indicative-readings

Indicative Readings

Science, meaning and human conversation in one source-linked dataset, in 15 languages.

Indicative Readings is a multilingual dataset in 15 languages, with Turkish as its source language. For each of the 78 fields of UNESCO's ISCED-F 2013 classification of education and work — from physics and law to nursing, catering and transport —, for 48 frontier topics of today's science (from the Higgs boson and dark matter to gene editing, protein folding and language models) and for 66 laws of nature and science, it first states what the field knows, with a verified reference, and then reads it in an indicative way: not only what a thing is, but what it points to — the measure, fit, order and care it displays. Each reading ends with an open question and an invitation to reflect, and leaves the reader free.

Every record has Turkish and English text. The same 664 of 816 records, including all 78 field profiles and all 165 sourced records of the locked test split, also have full record text in Arabic, Bengali, German, Spanish, Persian, French, Hindi, Indonesian, Japanese, Portuguese (Brazilian), Russian, Urdu, Chinese (Simplified). The other 152 records currently have record text in Turkish and English only; their translated field names may still be present. Full coverage in the other 13 languages remains in progress.

The result is an unusually broad, carefully structured research set: scientific explanation, an indicative reading, a familiar analogy with its boundary, and a five-step invitation in each sourced record. It spans physics, psychology, social work, law and everyday trades. The texts start from what the field knows; the reader is invited to think, without pressure.

At a glance

Sourced readings	738: 624 field readings, 48 frontier topics, 66 laws of nature and science
Fields of knowledge and work	78 detailed fields of ISCED-F 2013, each with a profile
Reference metadata checked	2179/2179 (100%) (88% peer-reviewed; 2048 Crossref metadata checks, 131 official-page text checks)
Languages	15: Turkish (source) and English in all 816 records; the same 664 records in all 15; 152 records in Turkish and English only
Meaning check against the Turkish source	English: 816/816 records (100%), 671 corrected; the other 13 languages: 8632/8632 record versions (100%), 2488 corrected; none open
Analogies that state where they stop	738/738 sourced readings (100%), each also stating its limits
Evidence level on every sourced reading	397 established, 317 well supported, 21 normative standard, 2 contested, 1 open question
Structured depth	the field's account, an indicative reading, analogy and its limit, five steps, limits and references in a median of 545 English words per sourced record
Instruction pairs	10264 in 15 languages, built only from the records' own text
Derived tasks	45470 in five kinds (separate science from reading, state the limits, say where the analogy stops, five steps, cite), and 18188 chosen/rejected pairs
Golden evaluation set	2475 items: the 165 sourced test records in all 15 languages, each with a gold answer, a seven-check rubric and scorer.py
Splits	by field, so no field leaks across splits; the test split (183 records) is locked by SHA-256 in manifest.json
Release gates	security and privacy scan of every file, withdrawal drill, recorded release decision

These 100% figures describe coverage of specified dataset checks. They are not model accuracy, hallucination or scientific-validity scores.

Why this dataset stands out

• Source first. Every sourced reading carries references whose bibliographic metadata were checked before release; numbers that could not be checked against a source were left out. Whether each explanation fully reflects its cited work remains the project's responsibility.
• Information and meaning, kept apart. Each record holds the field's account (what a thing is) and the indicative reading (what it points to) in separate fields. This gives researchers a concrete format for teaching and testing the move from fact to meaning without treating an interpretation as a scientific finding.
• Analogies with their limits. Every analogy says where it stops, and every reading says what it does not claim.
• A human approach. Five steps move from shared observation through evidence and a question to an invitation. The ending leaves room for the reader's judgment. Health, law and finance readings carry advice limits; health and care readings name who makes decisions.
• Parallel where complete. Turkish is the source. The same 664 records carry all 15 languages, and each target-language version was checked against the source field by field for meaning (who does what, negation, conditions, scope, certainty, idioms). Select languages to avoid treating the other 152 records as translated.
• Breadth with a common method. The Higgs boson, CRISPR and language models sit next to psychology, social work, law, nursing, catering, accounting and transport. The same source, analogy and limit structure makes those fields comparable.
• Built to be audited. Content hashes, a verification ledger, a file manifest, and public corrections and withdrawals.

A research strategy for future models

The indicative (mânâ-yı harfî) reading treats a fact as something that can point beyond itself. Keeping the scientific account, the interpretive reading and the analogy's stopping point in separate fields offers a practical way to train and evaluate meaning modelling, source-grounded explanation and respectful dialogue. The 10264 instruction pairs provide two tasks per checked record and language; the 45470 derived tasks and 18188 preference pairs train the separate moves one by one, and the golden set measures them on the locked test split.

This is a promising design hypothesis, not a demonstrated model advantage. No generative model trained on this dataset has been evaluated. Claims about hallucination rate, reinforcement fit, character efficiency or superiority over other models need controlled, reproducible model tests before they can be reported as results.

Quick start

from datasets import load_dataset

readings = load_dataset("medresetuzzehra/indicative-readings", "readings")["train"]   # one file per config; the split is a column
train = readings.filter(lambda r: r["split"] == "train")
r = train[0]
print(r["topic_en"], r["science_en"], r["reading_en"], r["analogy_limit_en"], sep="\n\n")
test = readings.filter(lambda r: r["split"] == "test")
print(test[0]["languages"], test[0]["reading_ar"], test[0]["reading_ja"], sep="\n\n")   # the test split in every language

pairs = load_dataset("medresetuzzehra/indicative-readings", "training_pairs")["train"]
tasks = load_dataset("medresetuzzehra/indicative-readings", "tasks")["train"].filter(lambda t: t["split"] == "train")   # never train on the test split

golden = load_dataset("medresetuzzehra/indicative-readings", "golden")["train"]
# Write one {"id": ..., "output": ...} line per golden item to answers.jsonl, then:
#   python scorer.py data/golden.jsonl answers.jsonl

Example

Freshwater’s density maximum near 4 °C and floating ice (Physics)

At ordinary pressure, pure freshwater has its greatest density near 4 °C and becomes less dense as it cools below that temperature. On freezing, an open crystal structure associated with hydrogen bonding forms; volume increases by about 9 percent, density falls and ice floats on liquid water. In sufficiently deep freshwater lakes in winter, cooling upper water can mix downward while it becomes denser; colder, lighter water may remain above and form ice. An ice cover slows heat loss from the water below and can help liquid water persist. Bottom water need not be exactly 4 °C, and fish survival is not guaranteed; depth, climate, dissolved oxygen and other conditions affect the outcome.

Reading. A harfî view may read floating ice and liquid water beneath it in some sufficiently deep freshwater lakes as a sign inviting reflection on the Giver of these possibilities of water. The density maximum near 4 °C and the lower density of ice are physical properties; an ice cover can slow heat loss. A water molecule does not consciously choose to protect fish, and fish are not guaranteed to survive winter in every lake. Keeping physics and lake conditions in view, we may reflect harfîly on this relation; the observation alone does not scientifically prove an agent.

1. In a cold region, a frozen freshwater lake can have ice above and liquid water below; we can also observe ice floating in a glass of fresh water.
2. Physics shows that pure water at ordinary pressure is densest near 4 °C and expands in volume by about 9 percent on freezing. In some sufficiently deep freshwater lakes the upper water cools and mixes, while lighter cold water may freeze on top; an ice cover slows heat loss below.
3. What possibilities do floating ice and liquid water beneath it in some lakes suggest to you, and where are the limits of that thought when lake depth and oxygen vary?
4. A water molecule does not think of fish. Keeping water-density physics and differing winter lake conditions in view, we may read this relation as a harfî sign inviting reflection on the Giver of water’s possibilities; the data alone do not compel an agent.
5. If you like, let us put ice in a glass of fresh water and watch it float together; consider for yourself why this alone does not tell us how fish fare in every lake.

Analogy. A cover over an outdoor pool in winter can slow heat loss from the water; ice on top can likewise slow heat loss from water beneath in some lakes. Where it stops: People place a pool cover, whereas ice forms through water’s density and freezing properties. Neither cover guarantees that water beneath will not freeze or that organisms will survive. The analogy carries only the slowing of heat loss by a layer on top.

Limits. This reading does not replace water physics or limnology and does not claim proof. The density maximum near 4 °C describes pure freshwater at ordinary pressure; salinity and pressure can alter it. Ice cover is not perfect insulation: shallow lakes can freeze farther down, oxygen beneath ice can fall, and fish kills can occur. We cannot assume every lake bottom is at 4 °C or every fish safely survives winter.

Contents

Config	Records	What it holds
readings	624	eight readings per field: the field's account, the indicative reading, an analogy and its limit, five steps, limits, references
frontiers	48	frontier topics of today's science, one reading each, built the same way as the readings
laws	66	laws of nature and science with their conditions of validity, read the same way
profiles	78	for each field, its answer to 'What is this universe?' and the reality the field rests on
fields	117	the 78 ISCED-F 2013 detailed fields covered and their broad and narrow groups, in English and Turkish
methods	20	the reading methods used, in plain words
references	2179	the works the readings rest on, each verified (see Quality)
training_pairs	10264	instruction and answer pairs built from the records' own text
tasks	45470	five derived tasks per sourced record and language: separate science from reading, state the limits, say where the analogy stops, write the five steps, cite the references
preferences	18188	chosen and rejected answers: the rejected one leaves out the limits or presents the reading as a proven result
golden	2475	the golden evaluation set: the locked test split in all 15 languages with gold reference answers, a mechanical rubric and the checks each item passed; score answers with scorer.py
coverage	78	which records cover which field

Every text field carries its language as a suffix: Turkish (_tr, the source language) and English (_en) in every record, and _ar, _bn, _de, _es, _fa, _fr, _hi, _id, _ja, _pt, _ru, _ur, _zh for the record text in the same 664 of 816 records (the other 152 have no target-language record text; translated field names may still appear). The languages column lists the languages a record carries; field_ columns give the field's name in every language. All language versions are the project's own texts. In the texts, harfî (Turkish, "like a letter") names the indicative view and ismî ("like a noun") the view of a thing for its own sake: a letter is read for what it points to, a noun for itself.

How the readings are made

1. Shared observation — something anyone can see or check.
2. The field's account — what the science knows, in plain words, with its evidence level and references.
3. Wonder as a question — how are this order, measure and fit possible?
4. An invitation to the inference — a work that needs such knowledge and measure makes one think of someone who knows and sets it.
5. Freedom — an open invitation to look again and think, without pressure.

Every analogy states where it stops, and every reading states what it does not claim: it does not change the science, it is not a proof, and it is not medical, legal or financial advice.

Quality

• References. 2179 of 2179 references were verified. For 2048 works, Crossref metadata matched the recorded title and year; for 131 official texts or repository items, the expected wording was found on the official page. The method for each work is in data/references.jsonl; the list is in SOURCES.md.
• Text rules. Every text was checked mechanically in its own language: no religious names, no argument from gaps, no pressure, the same numbers as the source.
• Meaning. Every English text was checked against the Turkish for meaning: who does what, negation, conditions, scope, certainty, time, literal and figurative speech. 816 records were checked; 671 were corrected and none is open. Every version in the other 13 languages was read a second time against the Turkish and the English by a project reader who did not write it, on the same points and on terminology, numbers and the language's own rule; 2488 of 8632 were corrected. Where a reader found the English ambiguous, the English was clarified and every language was read again on that point.
• Ledger and manifest. verification.jsonl lists every record with its content hash and checks; manifest.json lists every file with its hash.

Golden evaluation set and scorer

golden holds the 165 sourced records of the locked test split in all 15 languages (2475 items). Each item gives the prompt, the gold reference answer, the references, the rubric and the checks the record passed (public text rule, translation audit, meaning check, a reading of the science against its sources, verified references). scorer.py needs only the Python standard library and checks an answer mechanically on seven points: no religious name, no argument from gaps, no pressure, no claim that the reading is proven, the limits stated, the science kept apart from the reading, a free invitation.

The scorer was measured on the dataset's own texts: all 2475 gold answers and all 18188 chosen answers pass every check; it flags 9094 of 9094 rejected answers that present the reading as proven and 9094 of 9094 that leave out the limits. These figures test the scorer, not a model. The rubric is a floor: passing it does not show that an answer is true or well read, so the gold answers are there for comparison and human reading. Each language is scored with its own word lists and patterns, which scorer.py holds.

Intended use

Education and reflection; multilingual reading; training and evaluating language models to connect scientific knowledge with meaning honestly — source first, analogies with their limits, an open invitation instead of pressure.

Out of scope

Not a textbook; the scientific statements are short summaries with references. Not medical, legal or financial advice. The readings are ways of looking, not proofs, and must not be presented as anyone's quotations.

Splits

Records are split by field (all records of a field sit on one side): train, validation and test. The test split is locked: its record ids and hash are in manifest.json. It is a held-out part of this dataset, not an independent benchmark. tasks, preferences and training_pairs carry their record's split; the golden set is built from the test split only, so train on split == "train" rows.

Sources, license, corrections

• Sources and the origin of the method: SOURCES.md.
• License: Creative Commons Attribution 4.0 International (cc-by-4.0). Attribution: Medresetüzzehra project, Indicative Readings, version 3.3.0.
• Corrections and withdrawals: WITHDRAWALS.md; changes between versions: CHANGELOG.md. Please report problems in this repository's Discussions.

@misc{indicative_readings_3_3_0,
  title  = {Indicative Readings: every field of knowledge and the laws of nature read as signs},
  author = {{Medresetüzzehra project}},
  year   = {2026},
  note   = {Version 3.3.0, Creative Commons Attribution 4.0 International},
  howpublished = {\url{https://huggingface.co/datasets/medresetuzzehra/indicative-readings}}
}

Türkçe

Harfî Okumalar, kaynak dili Türkçe olan 15 dilli bir veri setidir. UNESCO'nun ISCED-F 2013 sınıflamasındaki 78 alanın her biri, 66 tabiat ve ilim kanunu ve bugünün ilminin 48 ileri konusu için önce ilmin bildiğini doğrulanmış kaynağıyla verir, sonra onu harfî (işaret eden) bakışla okur: bir şeyin yalnız ne olduğuna değil, neyi gösterdiğine — taşıdığı ölçüye, uyuma, düzene ve özene — bakar. Her okuma açık bir soru ve düşünmeye davetle biter; okuru serbest bırakır. Metinlerde dinî kelime kullanılmaz; vaaz edilmez ve ‘bilim açıklayamıyor’ diye delil kurulmaz.

Bir bakışta:

• 738 kaynaklı okuma;
• 2179/2179 kaynak künyesi denetlendi (%100); yüzde 88'i hakemli;
• 816 kaydın tamamı Türkçe ve İngilizce; aynı 664 kayıt 15 dilde (العربية, বাংলা, Deutsch, Español, فارسی, Français, हिन्दी, Bahasa Indonesia, 日本語, Português (Brasil), Русский, اردو, 中文（简体）), kalan 152 kayıt şimdilik yalnız Türkçe ve İngilizce;
• 816/816 kaydın İngilizcesi ve diğer dillerdeki 8632 kayıt sürümü Türkçeye karşı anlam yönünden denetlendi (%100); düzeltilen 671 ve 2488, açık sorun yok;
• 738/738 kaynaklı okumada benzetmenin sınırı açıkça yazılı (%100);
• 10264 eğitim çifti, 45470 türetilmiş görev ve 18188 tercih çifti, her biri kaydın dillerinde;
• 2475 maddelik altın değerlendirme seti (kilitli test bölmesi, 15 dil) ve standart kütüphaneyle çalışan scorer.py;
• alan bazlı, sızıntısız ve kilitli bölmeler.

Araştırma yönü. Bu düzen, ilmin anlattığını harfî anlam okumasından ve temsilin sınırından ayırır; ortak gözlemden başlayan beş adımlı hitap muhatabı serbest bırakır. Psikoloji, sosyal hizmet ve hukuk da fizik ve biyolojiyle aynı yapı içinde yer alır. Bilgiyi anlamla kaynağa bağlı kalarak birleştiren modeller için güçlü bir araştırma tasarımıdır. Yukarıdaki %100 oranlar veri kayıtlarının belirli denetimleri kapsama oranıdır; model başarımı değildir. Henüz bu veriyle eğitilmiş bir model ölçülmedi; halüsinasyon, pekiştirme uyumu veya karakter verimliliği için sonuç iddia edilmez.

Kaynaklar ve usulün kaynağı SOURCES.md dosyasındadır. Lisans CC BY 4.0'dır.