Static SELFIES embedding (320d) trained on COCONUT + ChEMBL34 pairs

This is a sentence-transformers model trained on the coconut-chembl34-mol-sim dataset. It maps inputs to a 320-dimensional dense vector space and can be used for semantic textual similarity, semantic search, paraphrase mining, classification, clustering, and more.

Model Details

Model Description

• Model Type: Sentence Transformer

• Maximum Sequence Length: 200 tokens
• Output Dimensionality: 320 dimensions
• Similarity Function: Cosine Similarity
• Supported Modality: Text
• Training Dataset:
  • coconut-chembl34-mol-sim
• Language: en
• License: cc-by-nc-sa-4.0

Model Sources

• Documentation: Sentence Transformers Documentation
• Repository: Sentence Transformers on GitHub
• Hugging Face: Sentence Transformers on Hugging Face

Full Model Architecture

SentenceTransformer(
  (0): StaticEmbedding({})
)

Usage

Direct Usage (Sentence Transformers)

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("tomaarsen/static-chemselfies-molsim")
# Run inference
sentences = [
    '[O] [=S] [Branch1] [Ring1] [O] [C] [S] [S] [S] [S] [S] [N]',
    '[S] [C] [=C] [C] [=C] [Ring1] [Branch1] [C] [=C]',
    '[O] [=C] [Branch1] [O] [C] [C] [=C] [C] [#C] [C] [#C] [C] [=C] [C] [C] [C]',
]
embeddings = model.encode(sentences)
print(embeddings.shape)
# [3, 320]

# Get the similarity scores for the embeddings
similarities = model.similarity(embeddings, embeddings)
print(similarities)
# tensor([[1.0000, 0.6880, 0.1453],
#         [0.6880, 1.0000, 0.1629],
#         [0.1453, 0.1629, 1.0000]])

Evaluation

Metrics

Semantic Similarity

• Datasets: molsim and molsim-fulltest
• Evaluated with EmbeddingSimilarityEvaluator

Baseline comparison

All numbers are Spearman/Pearson of embedding cosine similarity against MACCS fingerprint similarity labels on held-out SELFIES pairs of gbyuvd/coconut-chembl34-mol-sim.

* as reported by the authors of gbyuvd/chemembed-chemselfies on their held-out "combined-test" split of the same data family, with EmbeddingSimilarityEvaluator-style Spearman/Pearson of cosine vs MACCS. This model was evaluated on the dataset's own full test split (118,318 pairs), so the numbers are directly comparable in kind but not bit-identical in split.

The ~0.20 Spearman gap is the price of going static: a mean-pooled embedding table (one vector per SELFIES symbol, no context, ~1M parameters) cannot model token interactions the way the 11.1M-parameter BERT can. In exchange, this model is orders of magnitude faster at inference (a lookup + average; ~20x faster than a comparable-size encoder on GPU and ~80x on CPU) and supports Matryoshka truncation to 160/80 dims.

Limitations & Notes

• Static model: mean-pooled SELFIES token embeddings, no context or attention. It trails contextual encoders on fine-grained similarity; it trades that for ~1M parameters and very fast inference.
• Input format: SELFIES strings must be space-separated tokens, e.g. [C] [=C] [N] (use " ".join(re.findall(r"\[[^\]]*\]", s)) on raw SELFIES output). Raw glued SELFIES ([C][=C][N]) encodes to [UNK].
• The tokenizer's [CLS]/[SEP] post-processing was stripped for training and inference (noise for a token-bag model). Vocab: 3,095 SELFIES symbols, taken from gbyuvd/chemembed-chemselfies.
• Labels are MACCS fingerprint cosine similarity, so the embeddings primarily encode MACCS-like similarity. 1 epoch reached 0.7473 Spearman, 3 epochs 0.7519 — diminishing returns. Recipe: CosineSimilarityLoss + MatryoshkaLoss [320/160/80], lr 0.05, batch 2048, cosine scheduler, best-checkpoint restore; ~15 min training on a T4.
• License cc-by-nc-sa-4.0, inherited from the baseline model and dataset by gbyuvd.

Training Details

Training Dataset

coconut-chembl34-mol-sim

• Dataset: coconut-chembl34-mol-sim
• Size: 1,049,286 training samples
• Columns: sentence1, sentence2, and label
• Approximate statistics based on the first 100 samples:

  	sentence1	sentence2	label
  type	string	string	float
  modality	text	text
  details	min: 60 charactersmean: 436.88 charactersmax: 1766 characters	min: 48 charactersmean: 446.99 charactersmax: 1875 characters	min: 0.2mean: 0.54max: 0.92

• Samples:

  sentence1	sentence2	label
  [O] [=C] [N] [C] [C] [C] [C] [Ring1] [=Branch1] [=C] [O] [C] [Branch2] [Ring1] [Branch1] [O] [C] [O] [C] [Branch1] [Ring1] [C] [O] [C] [Branch1] [C] [O] [C] [Branch1] [C] [O] [C] [Ring1] [#Branch2] [O] [C] [Ring2] [Ring1] [C] [C] [C]	[O] [=C] [C] [=C] [Branch2] [Ring2] [=Branch1] [O] [C] [=Branch1] [Branch1] [=C] [Ring1] [=Branch1] [O] [C] [NH1] [C] [=Branch1] [=N] [=N] [C] [=Ring1] [Branch1] [C] [=C] [C] [=C] [C] [=C] [Ring1] [=Branch1] [C] [=C] [C] [=C] [Branch1] [C] [O] [C] [=C] [Ring1] [#Branch1] [O] [C]	0.57
  [O] [=C] [O] [C] [Branch1] [C] [C] [C] [Branch2] [Ring2] [O] [N] [=C] [Branch1] [C] [O] [C] [Branch2] [Ring1] [Branch1] [N] [=C] [Branch1] [C] [O] [C] [Branch1] [Ring1] [O] [C] [C] [O] [S] [=Branch1] [C] [=O] [=Branch1] [C] [=O] [O] [C] [C] [C] [=C] [C] [=C] [Branch1] [C] [O] [C] [=C] [Ring1] [#Branch1] [C] [=Branch2] [=Branch1] [#Branch2] [=N] [C] [Branch2] [=Branch1] [Ring2] [C] [=Branch2] [Branch1] [#C] [=N] [C] [C] [=Branch1] [C] [=O] [N] [Branch1] [=Branch2] [C] [Branch1] [C] [O] [C] [C] [Ring1] [Branch2] [C] [Branch2] [Ring2] [Ring2] [C] [=Branch1] [C] [=O] [N] [Branch1] [C] [C] [C] [Branch1] [S] [C] [=Branch1] [N] [=N] [C] [Ring2] [Branch1] [Ring2] [C] [Branch1] [C] [C] [C] [C] [O] [C] [C] [=C] [C] [=C] [C] [=C] [Ring1] [=Branch1] [C] [Branch1] [C] [C] [C] [C] [O] [C] [O]	[O] [=C] [C] [C] [C] [Ring1] [Ring2] [C] [C] [C] [C] [C] [=C] [C] [C] [=C] [C] [C] [C] [C] [C]	0.38
  [O] [=C] [N] [C] [Branch2] [Ring1] [C] [O] [C] [=Branch1] [C] [=O] [C] [C] [Branch1] [Ring2] [N] [C] [=O] [C] [Branch1] [C] [C] [C] [C] [Branch1] [C] [C] [C]	[O] [=C] [Branch3] [Ring1] [C] [Branch2] [O] [C] [C] [Branch1] [C] [O] [C] [Branch1] [C] [O] [C] [Branch1] [=Branch1] [O] [C] [Ring1] [Branch2] [C] [O] [C] [Branch1] [Ring1] [C] [=C] [Branch1] [C] [C] [C] [C] [C] [=C] [Branch2] [S] [Branch1] [C] [=Branch1] [C] [=O] [O] [C] [C] [Branch1] [C] [O] [C] [Branch1] [C] [O] [C] [Branch1] [=Branch1] [O] [C] [Ring1] [Branch2] [C] [O] [C] [Branch1] [Ring1] [C] [=C] [Branch1] [C] [C] [C] [C] [C] [=C] [Branch2] [=N] [=C] [C] [=Branch1] [C] [=O] [O] [C] [C] [C] [Branch2] [Branch1] [O] [C] [=Branch1] [C] [=O] [O] [C] [O] [C] [Branch1] [Ring1] [C] [O] [C] [Branch1] [C] [O] [C] [Branch1] [C] [O] [C] [Ring1] [#Branch2] [O] [C] [O] [C] [Branch1] [C] [C] [C] [Branch1] [P] [O] [C] [O] [C] [C] [Branch1] [C] [O] [C] [Branch1] [C] [O] [C] [Ring1] [Branch2] [O] [C] [Branch1] [C] [O] [C] [Ring2] [Ring1] [C] [O] [C] [Branch1] [C] [O] [C] [C] [Branch2] [Branch2] [#C] [C] [=Branch2] [#Branch1] [=C] [=C] [C] [C] [C] [Branch1] [C] [C] [C] [C] [C] [Branch2] [Branch1]...	0.49

• Loss: MatryoshkaLoss with these parameters:

  {
      "loss": "CosineSimilarityLoss",
      "matryoshka_dims": [
          320,
          160,
          80
      ],
      "matryoshka_weights": [
          1,
          1,
          1
      ],
      "n_dims_per_step": -1
  }
  

Evaluation Dataset

coconut-chembl34-mol-sim

• Dataset: coconut-chembl34-mol-sim
• Size: 10,000 evaluation samples
• Columns: sentence1, sentence2, and label
• Approximate statistics based on the first 100 samples:

  	sentence1	sentence2	label
  type	string	string	float
  modality	text	text
  details	min: 47 charactersmean: 424.23 charactersmax: 1441 characters	min: 37 charactersmean: 459.26 charactersmax: 1583 characters	min: 0.0mean: 0.53max: 0.92

• Samples:

  sentence1	sentence2	label
  [O] [=C] [C] [Branch2] [Ring2] [=Branch2] [O] [C] [O] [C] [Branch2] [Ring1] [Branch1] [C] [O] [C] [O] [C] [Branch1] [C] [C] [C] [Branch1] [C] [O] [C] [Branch1] [C] [O] [C] [Ring1] [=Branch2] [O] [C] [Branch1] [C] [O] [C] [Branch1] [C] [O] [C] [Ring2] [Ring1] [Ring2] [O] [=C] [Branch2] [Branch1] [O] [O] [C] [=C] [C] [Branch2] [Ring2] [Branch2] [O] [C] [O] [C] [Branch1] [Ring1] [C] [O] [C] [Branch1] [C] [O] [C] [Branch1] [C] [O] [C] [Ring1] [#Branch2] [O] [C] [O] [C] [Branch1] [Ring1] [C] [O] [C] [Branch1] [C] [O] [C] [Branch1] [C] [O] [C] [Ring1] [#Branch2] [O] [=C] [C] [Branch1] [C] [O] [=C] [Ring2] [Branch1] [Branch2] [Ring2] [Ring1] [=C] [C] [=C] [C] [=C] [Branch1] [C] [O] [C] [Branch1] [C] [O] [=C] [Ring1] [Branch2]	[S] [C] [C] [N] [C] [C] [C] [N] [C] [C] [Ring1] [#Branch2]	0.33
  [F] [C] [=Branch2] [Ring1] [C] [=C] [Branch1] [C] [F] [C] [C] [S] [C] [=N] [N] [=C] [Branch1] [Ring2] [S] [Ring1] [Branch1] [N] [F]	[O] [=C] [Branch1] [#Branch2] [N] [C] [C] [S] [C] [=C] [C] [=Ring1] [Branch1] [C] [C] [C] [N] [Branch1] [C] [C] [C] [C] [C] [N] [Branch2] [Ring1] [Branch1] [C] [C] [=C] [C] [=C] [C] [Branch1] [#Branch2] [O] [C] [Branch1] [C] [F] [Branch1] [C] [F] [F] [=C] [Ring1] [O] [C] [Ring1] [P] [C] [=Branch1] [C] [=O] [N] [C] [Ring2] [Ring1] [Branch2]	0.49
  [O] [=C] [Branch2] [=Branch1] [N] [O] [C] [C] [O] [C] [Branch2] [Branch1] [#Branch1] [O] [C] [=C] [C] [=C] [Branch2] [Ring1] [Branch1] [O] [C] [O] [C] [Branch1] [Ring1] [C] [O] [C] [Branch1] [C] [O] [C] [Branch1] [C] [O] [C] [Ring1] [#Branch2] [O] [C] [=C] [Branch1] [C] [O] [C] [=C] [Ring2] [Ring1] [Ring1] [OH1+1] [C] [Ring2] [Ring1] [#Branch1] [C] [=C] [C] [=C] [Branch1] [C] [O] [C] [=C] [Ring1] [#Branch1] [C] [Branch1] [C] [O] [C] [Branch1] [C] [O] [C] [Ring2] [Ring2] [#Branch1] [O] [C] [=C] [C] [=C] [C] [=C] [Branch1] [C] [O] [C] [Branch1] [Ring1] [O] [C] [=C] [Ring1] [=Branch2]	[O] [C] [Branch1] [C] [C] [C] [C] [C] [C] [C] [Ring1] [Ring1] [C] [C] [Ring1] [=Branch2] [C] [C] [C] [Ring1] [Branch1] [Branch1] [C] [O] [C]	0.42

• Loss: MatryoshkaLoss with these parameters:

  {
      "loss": "CosineSimilarityLoss",
      "matryoshka_dims": [
          320,
          160,
          80
      ],
      "matryoshka_weights": [
          1,
          1,
          1
      ],
      "n_dims_per_step": -1
  }
  

Training Hyperparameters

Non-Default Hyperparameters

• per_device_train_batch_size: 2048
• learning_rate: 0.05
• lr_scheduler_type: cosine
• warmup_steps: 0.1
• bf16: True
• per_device_eval_batch_size: 2048
• load_best_model_at_end: True
• seed: 12

All Hyperparameters

• per_device_train_batch_size: 2048
• num_train_epochs: 3
• max_steps: -1
• learning_rate: 0.05
• lr_scheduler_type: cosine
• lr_scheduler_kwargs: None
• warmup_steps: 0.1
• optim: adamw_torch_fused
• optim_args: None
• weight_decay: 0.0
• adam_beta1: 0.9
• adam_beta2: 0.999
• adam_epsilon: 1e-08
• optim_target_modules: None
• gradient_accumulation_steps: 1
• average_tokens_across_devices: True
• max_grad_norm: 1.0
• label_smoothing_factor: 0.0
• bf16: True
• fp16: False
• bf16_full_eval: False
• fp16_full_eval: False
• tf32: None
• gradient_checkpointing: False
• gradient_checkpointing_kwargs: None
• torch_compile: False
• torch_compile_backend: None
• torch_compile_mode: None
• use_liger_kernel: False
• liger_kernel_config: None
• use_cache: False
• neftune_noise_alpha: None
• torch_empty_cache_steps: None
• auto_find_batch_size: False
• log_on_each_node: True
• logging_nan_inf_filter: True
• include_num_input_tokens_seen: no
• log_level: passive
• log_level_replica: warning
• disable_tqdm: False
• project: huggingface
• trackio_space_id: None
• trackio_bucket_id: None
• trackio_static_space_id: None
• per_device_eval_batch_size: 2048
• prediction_loss_only: True
• eval_on_start: False
• eval_do_concat_batches: True
• eval_use_gather_object: False
• eval_accumulation_steps: None
• include_for_metrics: []
• batch_eval_metrics: False
• save_only_model: False
• save_on_each_node: False
• enable_jit_checkpoint: False
• push_to_hub: False
• hub_private_repo: None
• hub_model_id: None
• hub_strategy: every_save
• hub_always_push: False
• hub_revision: None
• load_best_model_at_end: True
• ignore_data_skip: False
• restore_callback_states_from_checkpoint: False
• full_determinism: False
• seed: 12
• data_seed: None
• use_cpu: False
• accelerator_config: {'split_batches': False, 'dispatch_batches': None, 'even_batches': True, 'use_seedable_sampler': True, 'non_blocking': False, 'gradient_accumulation_kwargs': None}
• parallelism_config: None
• dataloader_drop_last: False
• dataloader_num_workers: 0
• dataloader_pin_memory: True
• dataloader_persistent_workers: False
• dataloader_prefetch_factor: None
• dataloader_multiprocessing_context: None
• dataloader_in_order: True
• remove_unused_columns: True
• label_names: None
• train_sampling_strategy: random
• length_column_name: length
• ddp_find_unused_parameters: None
• ddp_bucket_cap_mb: None
• ddp_broadcast_buffers: False
• ddp_static_graph: None
• ddp_backend: None
• ddp_timeout: 1800
• fsdp: None
• fsdp_config: None
• deepspeed: None
• debug: []
• skip_memory_metrics: True
• do_predict: False
• resume_from_checkpoint: None
• local_rank: -1
• prompts: None
• batch_sampler: batch_sampler
• multi_dataset_batch_sampler: proportional
• router_mapping: {}
• learning_rate_mapping: {}
• warmup_ratio: None

Training Logs

• The bold row denotes the saved checkpoint.

Training Time

• Training: 14.4 minutes
• Evaluation: 1.2 minutes
• Total: 15.6 minutes

Framework Versions

• Python: 3.12.12
• Sentence Transformers: 6.0.1
• Transformers: 5.16.1
• PyTorch: 2.14.0+cu130
• Accelerate: 1.14.0
• Datasets: 5.0.1
• Tokenizers: 0.23.2

Additional Resources

• Training and Finetuning Embedding Models with Sentence Transformers: the end-to-end guide for training or finetuning Sentence Transformer models.
• Introduction to Matryoshka Embedding Models: variable-size embeddings that can be truncated with minimal quality loss.
• Binary and Scalar Embedding Quantization for Significantly Faster & Cheaper Retrieval: post-training compression of embedding vectors.
• Multimodal Embedding & Reranker Models with Sentence Transformers: use text, image, audio, and video models through the same API.
• Training and Finetuning Multimodal Embedding & Reranker Models with Sentence Transformers: train multimodal embedding models, with a Visual Document Retrieval walkthrough.

Citation

BibTeX

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",
    publisher = "Association for Computational Linguistics",
    url = "https://arxiv.org/abs/1908.10084",
}

MatryoshkaLoss

@misc{kusupati2024matryoshka,
    title={Matryoshka Representation Learning},
    author={Aditya Kusupati and Gantavya Bhatt and Aniket Rege and Matthew Wallingford and Aditya Sinha and Vivek Ramanujan and William Howard-Snyder and Kaifeng Chen and Sham Kakade and Prateek Jain and Ali Farhadi},
    year={2024},
    eprint={2205.13147},
    archivePrefix={arXiv},
    primaryClass={cs.LG}
}