LayaLink

LayaLink is a typed decisions for concept mapping and alignment. It is finetuned on convaiinnovations/laya.

Task Input (state) Questions and answers
ontology_alignment a source and a target concept (label, definition, parents, ontology) relation: equivalent, source_narrower (source is a subclass of target), source_broader, related, no_match, insufficient; exact_match: P(same meaning and scope)
entity_linking a text mention, its sentence, one candidate concept mapping: match, no_match, insufficient; accept: P(enough evidence to accept)
concept_selection a mention, its sentence, a list of candidate concepts concept: the matching candidate's id or none; resolvable: P(some candidate matches)

LayaLink judges candidates you supply; it does not search an ontology.

Installation

pip install "laya @ git+https://github.com/NandhaKishorM/laya.git@80a1da1fa5448ea12163f4d977935625a0aea670"

or

 pip install laya==0.4.34

This is the laya version LayaLink was trained and tested with (Python 3.10+).

Quickstart

import laya

agent = laya.Agent("sensein/layalink", device="cuda")      # or "mps" (Apple Silicon) / "cpu"
# a local copy works the same way: laya.Agent("path/to/LayaLink")

If loading hangs and TensorFlow is installed, set USE_TF=0 (as for the base model).

Ask the questions with exactly the wording below: it is the wording LayaLink was trained on. For concept_selection, build criteria from your candidates (id: "label (ONTOLOGY)", plus none).

Ontology alignment: how does SOURCE relate to TARGET? (equivalent)

A real row from the held-out oaei_anatomy test set (provenance: oaei_reference).

state = {'source': {'label': 'urinary bladder urothelium',
                    'definition': '',
                    'synonyms': ['transitional epithelium'],
                    'ontology': 'Adult Mouse Anatomy'},
         'target': {'label': 'Bladder Transitional Cell Epithelium',
                    'definition': '',
                    'synonyms': ['Bladder Urothelium', 'Bladder Transitional Epithelium'],
                    'parents': ['Urothelium'],
                    'ontology': 'NCI Thesaurus (anatomy)'}}

questions = {'relation': {'type': 'choice',
                          'instructions': 'Classify SOURCE relative to TARGET using the supplied '
                                          'definitions. Subclass direction matters. Association is not '
                                          'equivalence. Missing information means insufficient.',
                          'criteria': {'equivalent': 'Same class meaning and scope.',
                                       'source_narrower': 'Source is a proper subclass of target.',
                                       'source_broader': 'Target is a proper subclass of source.',
                                       'related': 'Explicit association but neither equivalence nor '
                                                  'subclass.',
                                       'no_match': 'Definitions establish different concepts; no mapping '
                                                   'is supported.',
                                       'insufficient': 'Available evidence cannot determine the '
                                                       'relation.'}},
             'exact_match': {'type': 'noul',
                             'instructions': 'Do the supplied definitions establish that source and target '
                                             'have the same class meaning and scope? False includes '
                                             'insufficient evidence.'}}

result = agent.predict(state, questions)

LayaLink's answer (from the evaluation run), with the gold label:

result['answers']['relation']  # -> 'equivalent'  (gold: 'equivalent')  {'equivalent': 0.9993, 'source_broader': 0.0002, 'source_narrower': 0.0001, 'related': 0.0001, 'no_match': 0.0001, 'insufficient': 0.0001}
result['answers']['exact_match']  # -> 'true'  (gold: 'true')  {'true': 0.9997, 'false': 0.0003}

Ontology alignment: how does SOURCE relate to TARGET? (source_narrower)

A real row from the held-out bioml2026 test set (provenance: derived_hierarchy).

state = {'source': {'label': 'Hyperlipoproteinemia, Type II',
                    'definition': 'An inheritable form of hyperlipidemia, in which there are excess lipids '
                                  'in the blood.',
                    'synonyms': ['Type II Hyperlipidemia'],
                    'parents': ['Hyperlipoproteinemia'],
                    'ontology': 'NCI Thesaurus'},
         'target': {'label': 'familial hyperlipidemia',
                    'definition': 'A lipid metabolism disease that is characterized by elevated levels of '
                                  'lipids in the blood.',
                    'synonyms': ['familial hyperlipoproteinemia', 'hyperlipemia', 'hyperlipidaemia'],
                    'parents': ['lipid metabolism disorder'],
                    'ontology': 'Disease Ontology'}}

questions = {'relation': {'type': 'choice',
                          'instructions': 'Classify SOURCE relative to TARGET using the supplied '
                                          'definitions. Subclass direction matters. Association is not '
                                          'equivalence. Missing information means insufficient.',
                          'criteria': {'equivalent': 'Same class meaning and scope.',
                                       'source_narrower': 'Source is a proper subclass of target.',
                                       'source_broader': 'Target is a proper subclass of source.',
                                       'related': 'Explicit association but neither equivalence nor '
                                                  'subclass.',
                                       'no_match': 'Definitions establish different concepts; no mapping '
                                                   'is supported.',
                                       'insufficient': 'Available evidence cannot determine the '
                                                       'relation.'}},
             'exact_match': {'type': 'noul',
                             'instructions': 'Do the supplied definitions establish that source and target '
                                             'have the same class meaning and scope? False includes '
                                             'insufficient evidence.'}}

result = agent.predict(state, questions)

LayaLink's answer (from the evaluation run), with the gold label:

result['answers']['relation']  # -> 'source_narrower'  (gold: 'source_narrower')  {'source_narrower': 0.9991, 'equivalent': 0.0002, 'source_broader': 0.0002, 'related': 0.0002, 'no_match': 0.0002, 'insufficient': 0.0002}
result['answers']['exact_match']  # -> 'false'  (gold: 'false')  {'false': 0.9997, 'true': 0.0003}

Entity linking: does this candidate match the mention in context?

A real row from the held-out ncbi_corpora test set (provenance: corpus_gold).

state = {'mention': 'metrizoate',
         'context': 'Frequencies of ventricular fibrillation were significantly lower (p less than 0.05) '
                    'after iopentol (0%) and iohexol (3%) than after metrizoate (22%).',
         'candidate': {'label': 'Metrizoic Acid',
                       'definition': 'A diagnostic radiopaque that usually occurs as the sodium salt.',
                       'synonyms': ['Benzoic acid, 3-(acetylamino)-5-(acetylmethylamino)-2,4,6-triiodo-'],
                       'ontology': 'MeSH'}}

questions = {'mapping': {'type': 'choice',
                         'instructions': 'Does the candidate match the meaning of the mention in this '
                                         'context?',
                         'criteria': {'match': 'Candidate matches the contextual meaning.',
                                      'no_match': 'Candidate has a different meaning.',
                                      'insufficient': 'Context does not resolve the mention.'}},
             'accept': {'type': 'noul',
                        'instructions': 'Does the context provide enough evidence to accept this candidate '
                                        'as the mention meaning?'}}

result = agent.predict(state, questions)

LayaLink's answer (from the evaluation run), with the gold label:

result['answers']['mapping']  # -> 'match'  (gold: 'match')  {'match': 0.9999, 'no_match': 0.0000, 'insufficient': 0.0000}
result['answers']['accept']  # -> 'true'  (gold: 'true')  {'true': 0.9993, 'false': 0.0007}

Concept selection: which candidate (or none) matches the mention?

A real row from the held-out ner_bionlp/BC2GM test set (provenance: silver_exact).

state = {'mention': 'Insulin',
         'context': 'Insulin-stimulated glucose transport in adipocytes is mediated by the insulin '
                    'receptor.',
         'entity_type': 'gene or protein',
         'candidates': [{'id': 'pr:000050339',
                         'label': 'insulin hormone',
                         'ontology': 'PR',
                         'definition': 'A proinsulin that has had the propeptide (C-peptide) excised, '
                                       'leaving the...'},
                        {'id': 'pr:P01315',
                         'label': 'insulin (pig)',
                         'ontology': 'PR',
                         'definition': 'A protein that is a translation product of the INS gene in pig.'},
                        {'id': 'pr:P01326',
                         'label': 'insulin-2 (mouse)',
                         'ontology': 'PR',
                         'definition': 'An insulin-2 that is encoded in the genome of mouse.'},
                        {'id': 'pr:P15127',
                         'label': 'insulin receptor (rat)',
                         'ontology': 'PR',
                         'definition': 'An insulin receptor that is encoded in the genome of rat.'},
                        {'id': 'pr:000050340',
                         'label': 'insulin hormone (human)',
                         'ontology': 'PR',
                         'definition': 'An insulin hormone in human.'}]}

questions = {'concept': {'type': 'choice',
                         'instructions': 'Which candidate concept matches the meaning of the mention in '
                                         'this context? Choose none when no listed candidate matches.',
                         'criteria': {'pr:000050339': 'insulin hormone (PR)',
                                      'pr:P01315': 'insulin (pig) (PR)',
                                      'pr:P01326': 'insulin-2 (mouse) (PR)',
                                      'pr:P15127': 'insulin receptor (rat) (PR)',
                                      'pr:000050340': 'insulin hormone (human) (PR)',
                                      'none': 'None of the listed candidate concepts matches the '
                                              'mention.'}},
             'resolvable': {'type': 'noul',
                            'instructions': 'Does any listed candidate concept match the meaning of the '
                                            'mention in this context?'}}

result = agent.predict(state, questions)

LayaLink's answer (from the evaluation run), with the gold label:

result['answers']['concept']  # -> 'pr:000050339'  (gold: 'pr:000050339')  {'pr:000050339': 0.7554, 'none': 0.2191, 'pr:P01315': 0.0182, 'pr:P01326': 0.0025, 'pr:000050340': 0.0025, 'pr:P15127': 0.0024}
result['answers']['resolvable']  # -> 'true'  (gold: 'true')  {'true': 0.8884, 'false': 0.1116}

Many inputs at once

results = agent.predict_batch([state_1, state_2, state_3], questions, batch_size=64)   # one shared question set
for r in results:
    print(r["answers"])

Each answer has choice + probabilities (choice questions) or noul = P(yes) (yes/no questions), plus confidence.

Using with Router

from laya import Router

router =  Router(models={"english": "sensein/layalink", "multilingual": "sensein/layalink"}, max_loaded=1, device="cuda")
state = {'source': {'label': 'urinary bladder urothelium',
                    'definition': '',
                    'synonyms': ['transitional epithelium'],
                    'ontology': 'Adult Mouse Anatomy'},
         'target': {'label': 'Bladder Transitional Cell Epithelium',
                    'definition': '',
                    'synonyms': ['Bladder Urothelium', 'Bladder Transitional Epithelium'],
                    'parents': ['Urothelium'],
                    'ontology': 'NCI Thesaurus (anatomy)'}}

questions = {'relation': {'type': 'choice',
                          'instructions': 'Classify SOURCE relative to TARGET using the supplied '
                                          'definitions. Subclass direction matters. Association is not '
                                          'equivalence. Missing information means insufficient.',
                          'criteria': {'equivalent': 'Same class meaning and scope.',
                                       'source_narrower': 'Source is a proper subclass of target.',
                                       'source_broader': 'Target is a proper subclass of source.',
                                       'related': 'Explicit association but neither equivalence nor '
                                                  'subclass.',
                                       'no_match': 'Definitions establish different concepts; no mapping '
                                                   'is supported.',
                                       'insufficient': 'Available evidence cannot determine the '
                                                       'relation.'}},
             'exact_match': {'type': 'noul',
                             'instructions': 'Do the supplied definitions establish that source and target '
                                             'have the same class meaning and scope? False includes '
                                             'insufficient evidence.'}}
result = router.predict(state, questions)
print(result["routing"]["repo"])          # should print sensein/layalink

import json
print(json.dumps(result, indent=2))

Output

{
  "model": "laya-rl-agent",
  "answers": {
    "relation": {
      "type": "choice",
      "choice": "equivalent",
      "probabilities": {
        "equivalent": 0.9993,
        "source_narrower": 0.0001,
        "source_broader": 0.0002,
        "related": 0.0001,
        "no_match": 0.0001,
        "insufficient": 0.0001
      },
      "confidence": 0.9964,
      "answer_confidence": 0.9993,
      "action": {
        "act_probability": 1.0
      }
    },
    "exact_match": {
      "type": "noul",
      "noul": 0.9997,
      "confidence": 0.9997,
      "answer_confidence": 0.9997,
      "action": {
        "act_probability": 1.0
      }
    }
  },
  "usage": {
    "input_tokens": 371,
    "output_tokens": 0,
    "state_tokens": 106,
    "state_tokens_dropped": 0,
    "truncated": false,
    "truncated_questions": []
  },
  "routing": {
    "model": "multilingual",
    "repo": "sensein/layalink",
    "reason": "Latin script, language not identified and no non-English letters; using default (multilingual)",
    "detection": {
      "script": "latin",
      "script_profile": {
        "latin": 1.0
      },
      "language": null,
      "is_english": true,
      "language_undecided": true,
      "diacritic_rate": 0.0,
      "non_latin_fraction": 0.0,
      "mixed_segment": null
    },
    "workflow": null
  }
}

Results

All test sets pooled

Measured right after training on every data/test/* set together (each question counted once). "Calibrated" applies the temperatures fitted on the calibration split.

Probabilities Decisions Accuracy ECE Brier Mean confidence
raw 25254 0.9647 0.0339 0.0420 0.9552
calibrated 25254 0.9647 0.0751 0.0490 0.9053

Per test set

From evaluation/run_evaluation.sh (results/LayaLink-20261115-214357). Accuracy and macro-F1 over the labels of each question; ECE = expected calibration error (lower is better). ner_bionlp labels are silver (lexical/semantic evidence), the others gold.

Test set Task / question n Accuracy Macro-F1 ECE
bioml2026 ontology_alignment/exact_match 603 0.974 0.961 0.0256
bioml2026 ontology_alignment/relation 603 0.939 0.947 0.0534
mappings ontology_alignment/exact_match 224 0.951 0.944 0.0471
mappings ontology_alignment/relation 224 0.951 0.959 0.0436
ncbi_corpora entity_linking/accept 1043 0.932 0.931 0.0647
ncbi_corpora entity_linking/mapping 1043 0.937 0.955 0.0601
ner_bionlp/BC2GM concept_selection/concept 162 0.765 0.836 0.2149
ner_bionlp/BC2GM concept_selection/resolvable 162 0.895 0.894 0.0608
ner_bionlp/BC2GM entity_linking/accept 785 0.977 0.827 0.1213
ner_bionlp/BC2GM entity_linking/mapping 785 0.978 0.840 0.1136
ner_bionlp/BC4CHEMD concept_selection/concept 176 0.869 0.909 0.2005
ner_bionlp/BC4CHEMD concept_selection/resolvable 176 0.943 0.933 0.0731
ner_bionlp/BC4CHEMD entity_linking/accept 798 0.961 0.906 0.0819
ner_bionlp/BC4CHEMD entity_linking/mapping 798 0.959 0.900 0.0703
ner_bionlp/BC5CDR-chem concept_selection/concept 224 0.862 0.884 0.1794
ner_bionlp/BC5CDR-chem concept_selection/resolvable 224 0.964 0.925 0.0773
ner_bionlp/BC5CDR-chem entity_linking/accept 947 0.963 0.933 0.0845
ner_bionlp/BC5CDR-chem entity_linking/mapping 947 0.962 0.932 0.0766
ner_bionlp/BC5CDR-disease concept_selection/concept 178 0.972 0.972 0.1869
ner_bionlp/BC5CDR-disease concept_selection/resolvable 178 0.966 0.966 0.0751
ner_bionlp/BC5CDR-disease entity_linking/accept 820 0.988 0.970 0.0447
ner_bionlp/BC5CDR-disease entity_linking/mapping 820 0.992 0.979 0.0415
ner_bionlp/JNLPBA concept_selection/concept 221 0.869 0.880 0.2261
ner_bionlp/JNLPBA concept_selection/resolvable 221 0.928 0.924 0.0555
ner_bionlp/JNLPBA entity_linking/accept 1100 0.986 0.869 0.0909
ner_bionlp/JNLPBA entity_linking/mapping 1100 0.983 0.845 0.0801
ner_bionlp/NCBI-disease concept_selection/concept 163 0.902 0.910 0.2037
ner_bionlp/NCBI-disease concept_selection/resolvable 163 0.914 0.914 0.0851
ner_bionlp/NCBI-disease entity_linking/accept 746 0.984 0.956 0.0622
ner_bionlp/NCBI-disease entity_linking/mapping 746 0.980 0.945 0.0485
ner_bionlp/linnaeus concept_selection/concept 334 0.973 0.983 0.2420
ner_bionlp/linnaeus concept_selection/resolvable 334 0.997 0.997 0.1415
ner_bionlp/linnaeus entity_linking/accept 1612 0.999 0.995 0.0713
ner_bionlp/linnaeus entity_linking/mapping 1612 0.998 0.990 0.0624
ner_bionlp/s800 concept_selection/concept 317 0.937 0.943 0.2935
ner_bionlp/s800 concept_selection/resolvable 317 0.937 0.936 0.1093
ner_bionlp/s800 entity_linking/accept 1535 0.996 0.989 0.0940
ner_bionlp/s800 entity_linking/mapping 1535 0.997 0.991 0.0816
oaei_anatomy ontology_alignment/exact_match 561 0.959 0.943 0.0399
oaei_anatomy ontology_alignment/relation 561 0.868 0.886 0.1080
oaei_conference ontology_alignment/exact_match 78 0.808 0.808 0.1921
oaei_conference ontology_alignment/relation 78 0.756 0.756 0.2200
bioml2026-benchmark/ncit-doid ontology_alignment/exact_match 2400 0.973 0.969 0.0249
bioml2026-benchmark/ncit-doid ontology_alignment/relation 2400 0.942 0.965 0.0480

Bio-ML 2026 local ranking (OAEI)

Rank the 100 released candidates per source concept; public valid split, scored exactly like the track's scoring_kit/score_local.py. The test gold is hidden: official test numbers come only from the organisers (CodaBench).

System Pair Queries MRR Hits@1 Hits@5 Hits@10
LayaLink ncit-doid 634 0.9044 0.8675 0.9464 0.9606
fuzzy label baseline ncit-doid 634 0.8985 0.8659 0.9322 0.9527

The test sets are this project's own held-out splits (grouped so no concept, pair or sentence is shared with training), not official benchmarks, except the Bio-ML ranking above.

Data Sources

The LayaLink training dataset integrates ontology alignment, biomedical entity normalization, and named entity recognition (NER) resources to support both concept-to-concept alignment and mention-to-concept mapping.

Data Sources

The dataset combines the following resources:

Source Datasets Description
OAEI 2025 oaei_anatomy, oaei_conference Reference alignments from the Anatomy and Conference tracks
OAEI Bio-ML 2026 bioml2026 NCITโ€“DOID training reference alignments only
Biomappings mappings Curated biomedical concept mappings
NCBI Normalization Corpora ncbi_corpora Entity normalization annotations from NCBI-disease and BC5CDR
BioNLP-Corpus ner_bionlp/* Biomedical entity mentions from eight NER datasets, linked to candidate ontology concepts using lexical and semantic matching

The BioNLP-Corpus collection includes:

  • BC2GM โ€” Gene and protein mentions
  • BC4CHEMD โ€” Chemical and drug mentions
  • BC5CDR-chem โ€” Chemical mentions
  • BC5CDR-disease โ€” Disease mentions
  • JNLPBA โ€” Biomedical entity mentions
  • NCBI-disease โ€” Disease mentions
  • linnaeus โ€” Species mentions
  • s800 โ€” Species mentions

Training Infrastructure

LayaLink was fine-tuned using two NVIDIA H100 GPUs, leveraging multi-GPU acceleration for training on the combined ontology alignment and biomedical entity mapping datasets.

Component Configuration
GPU 2 ร— NVIDIA H100
Training Multi-GPU fine-tuning
Training data Ontology alignments and biomedical entity mappings
Data split 80% training / 10% calibration / 10% testing

Benchmark compliance: In accordance with the OAEI Bio-ML 2026 rules, MONDO and UMLS were not used.

Limitations

  • A probability is a model estimate, not proof of logical equivalence. Check the logical consequences before asserting owl:equivalentClass.

License

LayaLink is released under the Apache License 2.0 (Apache-2.0).

The license permits commercial use, modification, distribution, and redistribution, subject to its terms.

See the Apache License 2.0 for details.

Note: The Apache-2.0 license applies to the LayaLink model and original code. Third-party datasets, ontologies, and other resources retain their respective licenses and usage restrictions.

Downloads last month
46
Safetensors
Model size
0.4B params
Tensor type
F16
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for sensein/layalink

Finetuned
(253)
this model