Model Card for Trait2Vec

Trait2Vec is a language model to embed organismal trait descriptions in a way that preserves the structure induced by a semantic similarity metric (e.g. SimGIC). The model was trained on the Character Similarity Dataset. It is fine-tuned from sentence-transformers/all-mpnet-base-v2. Through qualitative data exploration we observe the cosine similarity between embeddings of raw trait description is proportional to the semantic similarity of their corresponding ontological representations.

Model Details

Model Description

• Developed by: Juan Garcia, Soumyashree Kar, Jim Balhoff, Hilmar Lapp
• Model type: Sentence Transformer
• Language(s) (NLP): English
• License: MIT
• Fine-tuned from model: sentence-transformers/all-mpnet-base-v2

Model Sources

• Repository: Imageomics/char-sim

Uses

Trait2Vec has been qualitatively evaluated in the ability to embed raw trait descriptions in a way that preserves the structure of an ontology. Accordingly, we expect it to produce an alternative computational representation of the traits of an organism.

Direct Use

It can be used to embed the textual trait descriptions associated with an organism.

Bias, Risks, and Limitations

This model is finetuned from sentence-transformers/all-mpnet-base-v2, therefore it inherits its corresponding biases and risks. The training dataset(Character Similarity Dataset) introduces the biases of the single similarity metric and ontology. This means the embedding inherits that metric’s inductive biases, coverage gaps, and evolving definitions. Biological conclusions may differ under alternative metrics (e.g., Resnik, Jaccard) or other phenotype ontologies.

Recommendations

Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.

How to Get Started with the Model

First install the Sentence Transformers library:

pip install -U sentence-transformers

Then you can load this model and run inference.

from sentence_transformers import SentenceTransformer

# Download from the 🤗 Hub
model = SentenceTransformer("imageomics/trait2vec")
# Run inference
sentences = [
    'Form of distal portion of anteroventral process of ectopterygoid: varyingly falcate',
    'Ventral ridge of the coracoid with a posterior process: absent',
    'carpals: fully or partially ossified',
]
embeddings = model.encode(sentences)
print(embeddings.shape)
# [3, 256]

# Get the similarity scores for the embeddings
similarities = model.similarity(embeddings, embeddings)
print(similarities.shape)
# [3, 3]

Training Details

Training Data

This model was trained on the Character Similarity Dataset.

• Size: 438,516 training samples
• Columns: sentence1, sentence2, and score
• Approximate statistics based on the first 1000 samples:

  	sentence1	sentence2	score
  type	string	string	float
  details	min: 9 tokens mean: 42.84 tokens max: 164 tokens	min: 4 tokens mean: 22.8 tokens max: 164 tokens	min: 0.0 mean: 0.1 max: 0.61

• Samples:

  sentence1	sentence2	score
  Gill raker shape between ceratobranchial 1 and ceratobranchials 2--4: Homomorphic	Extent of development of inferior lamella of lateral ethmoid: inferior lamella absent	0.014706667500582846
  Gill raker shape between ceratobranchial 1 and ceratobranchials 2--4: Homomorphic	Shape of anal-fin pterygiophore tips: tips of pterygiophores shaped like an arrow-head; axial series of pterygiophores providing the ventral margin of the anal-fin base a scalloped appearance	0.030538703023734296
  Gill raker shape between ceratobranchial 1 and ceratobranchials 2--4: Homomorphic	Suprapreopercle: present	0.3385057414877959

• Loss: CoSENTLoss with these parameters:

  {
      "scale": 20.0,
      "similarity_fct": "pairwise_cos_sim"
  }
  

Training Hyperparameters

Non-Default Hyperparameters

• eval_strategy: steps

• per_device_train_batch_size: 64

• per_device_eval_batch_size: 64

• learning_rate: 2e-05

• num_train_epochs: 10

• warmup_ratio: 1e-06

• Training regime: fp32

Evaluation

We tested Trait2Vec on a hold-out split of 20% of the Character Similarity Dataset. No descriptor overlap was ensured.

Testing Data, Factors & Metrics

Testing Data

• Size: 111,628 evaluation samples
• Columns: sentence1, sentence2, and score
• Approximate statistics based on the first 1000 samples:

  	sentence1	sentence2	score
  type	string	string	float
  details	min: 9 tokens mean: 17.19 tokens max: 24 tokens	min: 3 tokens mean: 21.97 tokens max: 143 tokens	min: 0.0 mean: 0.1 max: 0.86

• Samples:

  sentence1	sentence2	score
  Ventral humeral ridge: or not	Metacarpals, Metacarpal I, presence: absent	0.05558851078197206
  Ventral humeral ridge: or not	Metapterygoid–quadrate fenestra: absent	0.004860625129173212
  Ventral humeral ridge: or not	Dorsal and ventral borders of the maxillary articular process: straight or slightly curved ventrally	0.10380567059620477

• Loss: CoSENTLoss with these parameters:

  {
      "scale": 20.0,
      "similarity_fct": "pairwise_cos_sim"
  }
  

Metrics

Semantic Similarity:

• Datasets: pheno-dev and pheno-test
• Evaluated with EmbeddingSimilarityEvaluator

Results

Summary

Trait2Vec embeds organismal trait descriptors in a way that preserves some of the ranking structure induced by the similarity metric of the ontology.

Environmental Impact

Experiments were conducted using a private infrastructure, which has a carbon efficiency of 0.432 kgCO$_2$eq/kWh. A cumulative of 20 hours of computation was performed on hardware of type A100 PCIe 40/80GB (TDP of 250W).

Total emissions are estimated to be 2.16 kgCO$_2$eq of which 0 percents were directly offset.

Estimations were conducted using the MachineLearning Impact calculator presented in:

@article{lacoste2019quantifying,
  title={Quantifying the Carbon Emissions of Machine Learning},
  author={Lacoste, Alexandre and Luccioni, Alexandra and Schmidt, Victor and Dandres, Thomas},
  journal={arXiv preprint arXiv:1910.09700},
  year={2019}
}

Model Architecture and Objective

SentenceTransformer(
  (0): Transformer({'max_seq_length': 256, 'do_lower_case': False}) with Transformer model: MPNetModel 
  (1): Pooling({'word_embedding_dimension': 768, 'pooling_mode_cls_token': True, 'pooling_mode_mean_tokens': False, 'pooling_mode_max_tokens': False, 'pooling_mode_mean_sqrt_len_tokens': False, 'pooling_mode_weightedmean_tokens': False, 'pooling_mode_lasttoken': False, 'include_prompt': True})
  (2): Dense({'in_features': 768, 'out_features': 256, 'bias': True, 'activation_function': 'torch.nn.modules.activation.Tanh'})
)

• Loss: CoSENTLoss with these parameters:

  {
      "scale": 20.0,
      "similarity_fct": "pairwise_cos_sim"
  }
  

Software

• Python: 3.10.16
• Sentence Transformers: 3.3.1
• Transformers: 4.48.1
• PyTorch: 2.5.1.post303
• Accelerate: 1.3.0
• Datasets: 2.14.4
• Tokenizers: 0.21.0

Citation

If you use this model in your research, please cite both it and the source model & method from which it was fine-tuned:

Model

@software{trait2vec2025,
  author = {Juan Garcia and Soumyashree Kar and Jim Balhoff and Hilmar Lapp},
  doi = {10.57967/hf/6892},
  title = {Trait2Vec (Revision f39747b)},
  version = {1.0.0},
  year = {2025},
  url = {https://huggingface.co/imageomics/trait2vec}
}

Source Model & Method

Sentence Transformers

@inproceedings{reimers-2019-sentence-bert,
    title = "Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks",
    author = "Reimers, Nils and Gurevych, Iryna",
    booktitle = "Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing",
    month = "11",
    year = "2019",
    pages = "3982-3992",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/D19-1410/",
    doi = "10.18653/v1/D19-1410"
}

CoSENTLoss

@online{kexuefm-8847,
    title={CoSENT: A more efficient sentence vector scheme than Sentence-BERT},
    author={Su Jianlin},
    year={2022},
    month={Jan},
    url={https://kexue.fm/archives/8847},
}

Acknowledgements

This work was supported by the Imageomics Institute, which is funded by the US National Science Foundation's Harnessing the Data Revolution (HDR) program under Award #2118240 (Imageomics: A New Frontier of Biological Information Powered by Knowledge-Guided Machine Learning). Any opinions, findings and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the National Science Foundation.

Model Card Authors

Juan Garcia

Model Card Contact

Please open a Discussion on the Community Tab with any questions on the model.