adambuttrick/ror-set-point-matching-train-validation-test

ROR set point matching: train, validation and test data

This dataset is for the task of affiliation matching. An affiliation string comes from a research paper, for example "Dept. of Physics, University of Tokyo, Japan". The task is to find the set of ROR records that the string names. ROR is the Research Organization Registry. A string can name zero, one or more organisations.

The dataset contains these parts:

• labelled affiliation strings, with train, validation and test splits;
• evaluation sets from other sources;
• the training data of the models in the ROR set point matching collection: teacher judgments, synthetic strings and the ROR records.

All ROR identifiers refer to ROR data release v2.14 (6 October 2026).

Configs

Config	Splits (rows)	Content
affiliations (default)	train 9,106 · validation 1,947 · test 1,947	Labelled affiliation strings from Crossref and OpenAlex, about one third in non-Latin scripts
affiliations_crossref	train 2,104 · validation 449 · test 447	The Crossref strings of affiliations, with the same splits
affiliations_openalex	train 3,501 · validation 748 · test 751	The OpenAlex strings of affiliations, with the same splits
affiliations_openalex_non_latin	train 3,501 · validation 750 · test 749	The OpenAlex non-Latin strings of affiliations, with the same splits
affiliations_unseen_institutions	train 16,452 · validation 1,914 · test 1,957	The labelled training strings, divided by institution. The validation and test splits contain only strings for institutions that are not in the train split
openalex_sample	train 11,281 · test 40,000	Random samples of OpenAlex affiliation strings: the development and test samples of openalex-affiliation-matcher
datacite	test 3,000	Labelled affiliation strings from DataCite metadata. 71% of them are equal, after normalisation, to a name of their gold record
teacher_judgments	train 196,018 · validation 3,982	Affiliation strings with probabilities from a teacher model, one for each candidate ROR record
synthetic_non_latin	train 133,757	Synthetic affiliation strings in non-Latin scripts, made by a language model from ROR records
ror_records	train 143,284	All active and inactive ROR records, with a text form for models

Load a config with the datasets library:

from datasets import load_dataset
ds = load_dataset("adambuttrick/ror-set-point-matching-train-validation-test", "affiliations")

Fields

In each split, id is unique (ror_id in ror_records).

affiliations

Field	Type	Description
id	string	The row number in the split
text	string	The affiliation string
ror_ids	list of strings	The ROR records that the string names. An empty list means that the string names no organisation
source	string	crossref, openalex or openalex_non_latin
source_line	int	The line in the source file

affiliations_crossref, affiliations_openalex, affiliations_openalex_non_latin

The rows of affiliations with the source crossref, openalex or openalex_non_latin, unchanged and in the same order. id is the id of the row in affiliations, so it is not a row number in these configs.

affiliations_unseen_institutions

Field	Type	Description
id	string	<origin>:<row>. The origin is the affiliations train split (train) or one of the openalex_sample train samples (dev_v1, dev_v2)
text, ror_ids		As in affiliations
origin	string	train, dev_v1 or dev_v2
source	string	The source of the string
script	string	Latin or non-Latin. A string is non-Latin when fewer than half of its letters are Latin

The file data/affiliations_unseen_institutions/held_out_records.jsonl lists the 1,255 held-out institutions. Each row gives the ROR id (ror) and the stratum of the record (type, script, freq and stratum).

openalex_sample

Field	Type	Description
id, text, ror_ids		As in affiliations
sample	string	v1 or v2
confidence	string	high, medium or low: the confidence of the label

datacite

id, text and ror_ids, as in affiliations.

teacher_judgments

Field	Type	Description
id, text		As in affiliations
judgments	list	{ror_id, p}, where p is the probability, from the teacher, that the string names the record

synthetic_non_latin

Field	Type	Description
id	string	<batch>:<record key>:<n>: the batch of the generator, the key of the record in the batch, and the number of the string
text, ror_ids		As in affiliations
ror_id	string	The record that the string was made from
form	string	full (the full local name), unit (a department or laboratory, then the organisation), short (an abbreviation), mixed (a mixed or transliterated form) or other_script (a second script that is usual in the country)
confidence	string	high, or low when the generator did not know an established local name
language	string	The language and script that the generator wrote in

The models of the collection identify a synthetic string by its line number. Do not change the order of the rows.

ror_records

Field	Type	Description
ror_id	string	The ROR id
status	string	active or inactive
name	string	The display name
acronyms, aliases	list of strings
labels	list	{v, lang}: names in other languages
places	list	{city, region, country, cc}
types	list of strings	The ROR types
relations	object	parent, child, related, predecessor and successor, each a list of {id, name}
text	string	The form that the models read

text has this form, with a maximum of 8 names in each relation list:

name | acronyms | other names | places | types | status | parent: ...; child: ...; related: ...

The folder data/ror_records also holds two maps:

• withdrawn.json: for each withdrawn ROR record, its successor, or null. A withdrawn record goes to its successor when it has exactly one successor link; the chain continues through withdrawn successors and ends at an active or inactive record.
• openalex_to_ror.json: for each OpenAlex institution, its ROR id.

With these maps, the folder is a complete records folder for the training and matching code.

How the data was made

ror_records

One row for each active (141,387) and inactive (1,897) record of ROR data release v2.14. Of the 1,433 withdrawn records, 1,248 go to an active successor, 36 to an inactive successor, and 149 to none. openalex_to_ror.json comes from the institutions file of openalex-affiliation-matcher v3.0.0.

affiliations

• The strings come from three labelled files: Crossref (3,000), OpenAlex (5,000) and OpenAlex non-Latin (5,000).
• A seeded procedure divided them into 70% train, 15% validation and 15% test (seed 1363).
• Each split has the same mix of source and script.
• Strings that are equal after the split's normalisation (NFKC, collapsed whitespace, casefold) are always in the same split.
• Strings that are also in openalex_sample train are always in the train split.
• The split does not use the labels.
• affiliations_crossref, affiliations_openalex and affiliations_openalex_non_latin divide each split of affiliations by source. Thus each source has the same 70/15/15 split.

affiliations_unseen_institutions

This config measures how well a model finds institutions that it did not see in training. It was made as follows (seed 20261022):

1. Collect the labelled strings: affiliations train and openalex_sample train (20,387 strings). A label on a withdrawn record goes to its successor.
2. Put each ROR record of the labels (8,340) into a stratum: the type of the record, the script of its strings, and the number of its strings (1, 2–3, 4–9 or 10+).
3. In each stratum, hold out 15% of the records, in sha256 order of the seed and the ROR id. The count is rounded up or down at random, so that small strata do not always round down. This gives 1,255 held-out records.
4. Move every string that names a held-out record out of the train split. Also move 15% of the strings that name no record.
5. Put the moved strings into groups. Strings that name the same held-out record are in one group.
6. Give the groups, in sha256 order, to the validation and test splits. Each group goes to the split that has fewer strings at that time. Do this in each source and script.
7. Remove duplicates. Strings are compared after normalisation (NFKC, casefold, and non-word characters as spaces). A train string that is equal to a moved string leaves the train split. A moved string that is equal to an affiliations validation string, or to a teacher string, is dropped. 64 strings are in no split.

The train split has 6,519 records. The largest group (399 strings) holds every string that names CNRS (https://ror.org/02feahw73). It is in the validation split, so the test split has no CNRS strings.

openalex_sample

The train split contains the development samples of openalex-affiliation-matcher (v1 and v2). The test split contains its test samples v1 and v2. Each sample is a random draw of affiliation strings from OpenAlex.

teacher_judgments

The teacher judgments of openalex-affiliation-matcher v3.0.0 (train_v1_jev_labels.jsonl.gz of that release) come from Jev, a decision model from TypeSafe AI. The release has one row for each (string, OpenAlex institution) pair. In this config:

• an OpenAlex id goes to the ROR id of its institution, and a withdrawn ROR id goes to its successor;
• a judgment with no ROR id is dropped (68 have no ROR id, and 105 have a withdrawn record with no successor);
• when two judgments of one string get the same ROR id, the higher p stays (7 times).

The config has 3,763,451 judgments.

synthetic_non_latin

A language model (gpt-6.1-sol, through the codex CLI, with no tools) wrote affiliation strings in the local script for each active or inactive record whose first place is in one of 51 countries:

Script group	Countries	Strings
Han, kana and Hangul	CN, TW, HK, MO, JP, KR, KP	92,072
Cyrillic	RU, UA, BY, KZ, KG, TJ, MN, BG, RS, MK	16,705
Arabic script	IR, AF, EG, SA, AE, IQ, JO, LB, SY, PS, KW, QA, BH, OM, YE, LY, SD, DZ, MA, TN, MR	11,731
Other	TH, LA, KH, MM, GR, CY, IL, GE, AM, ET, BD, NP, LK	13,249

• The records went to the model in batches of 40 (32,605 records, 816 batches).
• For each record, the model wrote one string of each form full, unit, short and mixed (32,605 each), and an other_script string when a second script is usual in the country (3,337).
• The labels are the source record, plus each related record that the model says the string names (7,237 strings).
• 121,927 strings have high confidence and 11,830 low.

A blind check of 240 strings (60 from each script group) by judges of another model family found these rates, weighted by the share of each group: wrong organisation 0.3%, unnatural form 10.3%, wrong labels 3.9%. The label errors had two patterns. Two label rules then corrected all strings:

Rule	Strings	What changes	Count
1	The source record is not healthcare, and the string names the Russian Academy of Sciences	The Academy is added. A named Siberian, Ural or Far Eastern branch is also added	898 strings, 1,067 labels added
2	The source record is healthcare	Labels whose record is not healthcare are dropped	1,931 strings, 1,955 labels dropped

Labels

The annotation guidelines are in annotation_guidelines_v1.2.txt. Example rules: label a parent organisation when the string names it; do not label an organisation that is only a part of the name of a hospital; do not use withdrawn ROR records.

Config	Labels
affiliations, openalex_sample	As annotated. In affiliations, some labels can be withdrawn records
affiliations_unseen_institutions	As in affiliations, but a label on a withdrawn ROR record goes to its successor
synthetic_non_latin	The source record and the related records that the generator named, then the two label rules above

Training strings that the models do not use

The models of the collection apply this rule:

• A teacher string is not used when the teacher gives p ≥ 0.5 to a held-out record of affiliations_unseen_institutions (28,474 train strings).
• A teacher string is not used when it is equal, after normalisation, to an evaluation string (11 more).
• A labelled or synthetic string is not used when it is equal to an evaluation string.
• A synthetic string is not used when it names a held-out record (2,871 strings). Thus no training row of the released retriever and set model has a held-out record as a positive (a label, or a teacher p ≥ 0.5).

A teacher string that gives a held-out record a p below 0.5 stays. Stage 1 of the set model uses the teacher's p as its target. Thus 9,606 of its strings have a target from 0.1 to 0.49 for a held-out record.

The evaluation strings are the validation and test splits of affiliations_unseen_institutions, and the affiliations validation split.

Rules for use

• Do not train on a validation split or a test split.
• Make all choices on the validation splits. Score the test splits one time, at the end.
• The per-source configs hold the same strings as affiliations. Do not combine them with affiliations.
• affiliations_unseen_institutions uses the same strings as affiliations train and openalex_sample train. A model that trained on those two splits saw the held-out institutions. For such a model, do not use affiliations_unseen_institutions.
• teacher_judgments and synthetic_non_latin can contain strings that are equal to evaluation strings. Before you score a model, remove each evaluation string that is equal, after normalisation, to a training string.
• Score with exact match, and with micro precision and recall over ROR ids. Exact match is the share of strings whose answer is exactly the set of gold records.

Sources and licences

Part	Source	Licence
ROR records	ROR data release v2.14	CC0
OpenAlex strings and institutions	OpenAlex	CC0
Crossref strings	Crossref metadata	See Crossref's metadata terms
DataCite strings	DataCite metadata	CC0
Teacher judgments	Release v3.0.0 of openalex-affiliation-matcher	As in that release
Synthetic strings	Made by gpt-6.1-sol from ROR records	—

Models

Each model card gives the configs and splits that the model used. Each model repository contains the exact training rows and the outputs on the evaluation splits.

Model	Role
ror-set-point-retriever-e5-base	Finds candidate ROR records for a string
ror-set-point-set-model-mmbert-base	Selects the records that the string names
ror-set-point-pair-scorer-baseline	Baseline
Variants	ror-set-point-retriever-e5-base-no-synthetic, ror-set-point-retriever-embeddinggemma-300m, ror-set-point-set-model-bge-reranker-v2-m3, ror-set-point-set-model-embeddinggemma-300m

Results of the matcher on the test splits, in percent:

Test split	Strings	System	Exact match	Precision	Recall
affiliations	1,947	openalex-affiliation-matcher v3.0.0, local	78.4	94.3	79.0
		Baseline	86.1	96.1	88.1
		Retriever and set model	90.0	95.7	92.6
affiliations_crossref	447	openalex-affiliation-matcher v3.0.0, local	85.9	96.7	88.6
		Baseline	91.7	97.9	93.9
		Retriever and set model	92.8	97.3	95.5
affiliations_openalex	751	openalex-affiliation-matcher v3.0.0, local	79.6	94.9	81.7
		Baseline	87.5	96.4	90.3
		Retriever and set model	90.0	96.0	93.0
affiliations_openalex_non_latin	749	openalex-affiliation-matcher v3.0.0, local	72.6	91.0	68.3
		Baseline	81.4	94.0	80.8
		Retriever and set model	88.4	94.1	89.6
affiliations_unseen_institutions	1,957	openalex-affiliation-matcher v3.0.0, local	77.0	95.9	80.0
		Baseline	79.2	93.6	85.1
		Retriever and set model	83.6	94.6	86.9
datacite	3,000	openalex-affiliation-matcher v3.0.0, local	92.4	97.7	93.2
		Baseline	91.9	95.1	95.6
		Retriever and set model	93.5	98.0	94.5
openalex_sample, sample v1	20,000	openalex-affiliation-matcher v3.0.0, local	81.6	95.3	83.2
		Baseline	86.9	95.6	90.0
		Retriever and set model	89.9	96.1	92.1
openalex_sample, sample v2	20,000	openalex-affiliation-matcher v3.0.0, local	82.6	94.9	84.2
		Baseline	88.0	95.2	91.5
		Retriever and set model	90.4	95.7	93.3

The answers of the retriever and set model are in the repository of the set model. The answers of v3.0.0 and of the baseline are in the repository of the baseline (ror-set-point-pair-scorer-baseline). On affiliations_unseen_institutions, the baseline trained on the train split of that config. On the other splits, it trained on the labelled strings of all institutions. On the links to held-out institutions of affiliations_unseen_institutions test, recall is 89.6 for the retriever and set model and 86.3 for the baseline. On the validation split, recall is 78.7 and 84.9. The cause is CNRS: the set model often misses it when a string names it only with its acronym. Without the CNRS strings, recall is 86.5 and 83.5.