ChemMRL: SMILES Matryoshka Representation Learning Embedding Transformer

This is a Chem-MRL (sentence-transformers) model finetuned from Derify/ModChemBERT-IR-BASE on the pubchem_10m_genmol_similarity dataset. It maps SMILES to a 1024-dimensional dense vector space and can be used for molecular similarity, semantic search, database indexing, molecular classification, clustering, and more.

Model Details

Model Description

• Model Type: ChemMRL (Sentence Transformer)
• Base model: Derify/ModChemBERT-IR-BASE
• Maximum Sequence Length: 512 tokens
• Output Dimensionality: 1024 dimensions
• Similarity Function: Tanimoto
• Training Dataset:
  • pubchem_10m_genmol_similarity
• License: apache-2.0

Model Sources

• Repository: Chem-MRL on GitHub
• Demo App Repository: Chem-MRL-demo on GitHub

Full Model Architecture

SentenceTransformer(
  (0): Transformer({'max_seq_length': 512, 'do_lower_case': False, 'architecture': 'ModChemBertModel'})
  (1): Pooling({'word_embedding_dimension': 1024, 'pooling_mode_cls_token': False, 'pooling_mode_mean_tokens': True, '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): Normalize()
)

Usage

Direct Usage (Chem-MRL)

First install the Chem-MRL library:

pip install -U chem-mrl>=0.7.3
pip install -U "transformers>=4.56.1,<5.0.0"

Then you can load this model and run inference.

from chem_mrl import ChemMRL

# Download from the 🤗 Hub
model = ChemMRL(
    "Derify/ChemMRL",
    trust_remote_code=True,
    model_kwargs={"torch_dtype": "bfloat16"},
)
# Run inference
sentences = [
    'OCCCc1cc(F)cc(F)c1',
    'Fc1cc(F)cc(-n2cc[o+]n2)c1',
    'CCC(C)C(=O)C1(C(NN)C(C)C)CCCC1',
]
embeddings = model.backbone.encode(sentences)
print(embeddings.shape)
# [3, 1024]

# Get the similarity scores for the embeddings
similarities = model.backbone.similarity(embeddings, embeddings)
print(similarities)
# tensor([[1.0000, 0.3876, 0.0078],
#         [0.3876, 1.0000, 0.0028],
#         [0.0078, 0.0028, 1.0000]])

Direct Usage (Sentence Transformers)

Click to see the direct usage in 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(
    "Derify/ChemMRL",
    # SentenceTransformer doesn't support tanimoto similarity natively so we set a different similarity function here
    similarity_fn_name="cosine",
    trust_remote_code=True,
    model_kwargs={"torch_dtype": "bfloat16"},
)
# Run inference
sentences = [
    'OCCCc1cc(F)cc(F)c1',
    'Fc1cc(F)cc(-n2cc[o+]n2)c1',
    'CCC(C)C(=O)C1(C(NN)C(C)C)CCCC1',
]
embeddings = model.encode(sentences)
print(embeddings.shape)
# [3, 1024]

# Get the similarity scores for the embeddings
similarities = model.similarity(embeddings, embeddings)
print(similarities)
# tensor([[1.0000, 0.5587, 0.0155],
#         [0.5587, 1.0000, 0.0055],
#         [0.0155, 0.0055, 1.0000]])

Evaluation

Metrics

Semantic Similarity

• Dataset: pubchem_10m_genmol_similarity
• Evaluated with chem_mrl.evaluation.EmbeddingSimilarityEvaluator.EmbeddingSimilarityEvaluator with these parameters:

  {
      "precision": "float32"
  }
  

Training Details

Training Dataset

pubchem_10m_genmol_similarity

• Dataset: pubchem_10m_genmol_similarity at 9aec8fd
• Size: 19,381,001 training samples
• Columns: smiles_a, smiles_b, and label
• Approximate statistics based on the first 1000 samples:

  	smiles_a	smiles_b	label
  type	string	string	float
  details	min: 17 tokens mean: 42.36 tokens max: 122 tokens	min: 11 tokens mean: 40.93 tokens max: 122 tokens	min: 0.02 mean: 0.56 max: 1.0

• Samples:

  smiles_a	smiles_b	label
  COc1ccc(NC(=O)C2CC[NH+](C(C)C(=O)Nc3ccc(C(=O)Nc4ccc(F)c(F)c4)cc3C)CC2)cc1NC(=O)C1CCCCC1	Cc1cc(C(=O)Nc2ccc(F)c(F)c2)ccc1NC(=O)C(C)[NH+]1CCC(C(=O)Nc2cccc(NC(=O)C3CCCCC3)c2)CC1	0.8495575189590454
  OCCN1CC[NH+](Cc2ccccc2OC2CC2)CC1	OCCN1CC[NH+](Cc2ccccc2On2cccn2)CC1	0.6615384817123413
  CC1CN(C(=O)C2CC[NH+](Cc3cccc(C(N)=O)c3)CC2)CC(C)O1	CC1CN(C(=O)C2CC[NH+](Cc3ccccc3)CC2)CC(C)O1	0.7123287916183472

• Loss: Matryoshka2dLoss with these parameters:

  {
      "loss": "TanimotoSentLoss",
      "n_layers_per_step": -1,
      "last_layer_weight": 2.0,
      "prior_layers_weight": 1.0,
      "kl_div_weight": 0.0,
      "kl_temperature": 0.0,
      "matryoshka_dims": [
          1024,
          512,
          256,
          128,
          64,
          32,
          16,
          8
      ],
      "matryoshka_weights": [
          1,
          1,
          1,
          1,
          1,
          1,
          1,
          1
      ],
      "n_dims_per_step": -1
  }
  

Evaluation Dataset

pubchem_10m_genmol_similarity

• Dataset: pubchem_10m_genmol_similarity at 9aec8fd
• Size: 1,080,394 evaluation samples
• Columns: smiles_a, smiles_b, and label
• Approximate statistics based on the first 1000 samples:

  	smiles_a	smiles_b	label
  type	string	string	float
  details	min: 16 tokens mean: 42.05 tokens max: 101 tokens	min: 11 tokens mean: 40.23 tokens max: 104 tokens	min: 0.0 mean: 0.57 max: 1.0

• Samples:

  smiles_a	smiles_b	label
  N#CCCN(Cc1cnc(N)cn1)C1CC1	N#CCCN(Cc1cnc(N)cn1)C1CCCC1	0.8600000143051147
  N#CCCN(Cc1cnc(N)cn1)C1CC1	N#CCCN(Cc1cnc(N)cn1)C1CCOCC1	0.7962962985038757
  N#CCCN(Cc1cnc(N)cn1)C1CC1	N#CCCN(Cc1cnc(N)cn1)CC(F)F	0.5517241358757019

• Loss: Matryoshka2dLoss with these parameters:

  {
      "loss": "TanimotoSentLoss",
      "n_layers_per_step": -1,
      "last_layer_weight": 2.0,
      "prior_layers_weight": 1.0,
      "kl_div_weight": 0.0,
      "kl_temperature": 0.0,
      "matryoshka_dims": [
          1024,
          512,
          256,
          128,
          64,
          32,
          16,
          8
      ],
      "matryoshka_weights": [
          1,
          1,
          1,
          1,
          1,
          1,
          1,
          1
      ],
      "n_dims_per_step": -1
  }
  

Training Hyperparameters

Non-Default Hyperparameters

• eval_strategy: steps
• per_device_train_batch_size: 192
• per_device_eval_batch_size: 512
• learning_rate: 8e-06
• weight_decay: 1e-05
• max_grad_norm: None
• lr_scheduler_type: warmup_stable_decay
• lr_scheduler_kwargs: {'num_decay_steps': 100943, 'warmup_type': 'linear', 'decay_type': '1-sqrt'}
• warmup_steps: 100943
• data_seed: 42
• bf16: True
• bf16_full_eval: True
• tf32: True
• optim: stable_adamw
• optim_args: decouple_lr=True,max_lr=8.0e-6
• gradient_checkpointing: True
• eval_on_start: True

All Hyperparameters

Click to expand

• overwrite_output_dir: False
• do_predict: False
• eval_strategy: steps
• prediction_loss_only: True
• per_device_train_batch_size: 192
• per_device_eval_batch_size: 512
• per_gpu_train_batch_size: None
• per_gpu_eval_batch_size: None
• gradient_accumulation_steps: 1
• eval_accumulation_steps: None
• torch_empty_cache_steps: None
• learning_rate: 8e-06
• weight_decay: 1e-05
• adam_beta1: 0.9
• adam_beta2: 0.999
• adam_epsilon: 1e-08
• max_grad_norm: None
• num_train_epochs: 3
• max_steps: -1
• lr_scheduler_type: warmup_stable_decay
• lr_scheduler_kwargs: {'num_decay_steps': 100943, 'warmup_type': 'linear', 'decay_type': '1-sqrt'}
• warmup_ratio: 0.0
• warmup_steps: 100943
• log_level: passive
• log_level_replica: warning
• log_on_each_node: True
• logging_nan_inf_filter: True
• save_safetensors: True
• save_on_each_node: False
• save_only_model: False
• restore_callback_states_from_checkpoint: False
• no_cuda: False
• use_cpu: False
• use_mps_device: False
• seed: 42
• data_seed: 42
• jit_mode_eval: False
• bf16: True
• fp16: False
• fp16_opt_level: O1
• half_precision_backend: auto
• bf16_full_eval: True
• fp16_full_eval: False
• tf32: True
• local_rank: 0
• ddp_backend: None
• tpu_num_cores: None
• tpu_metrics_debug: False
• debug: []
• dataloader_drop_last: False
• dataloader_num_workers: 0
• dataloader_prefetch_factor: None
• past_index: -1
• disable_tqdm: False
• remove_unused_columns: True
• label_names: None
• load_best_model_at_end: False
• ignore_data_skip: False
• fsdp: []
• fsdp_min_num_params: 0
• fsdp_config: {'min_num_params': 0, 'xla': False, 'xla_fsdp_v2': False, 'xla_fsdp_grad_ckpt': False}
• fsdp_transformer_layer_cls_to_wrap: None
• 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
• deepspeed: None
• label_smoothing_factor: 0.0
• optim: stable_adamw
• optim_args: decouple_lr=True,max_lr=8.0e-6
• adafactor: False
• group_by_length: False
• length_column_name: length
• project: huggingface
• trackio_space_id: trackio
• ddp_find_unused_parameters: None
• ddp_bucket_cap_mb: None
• ddp_broadcast_buffers: False
• dataloader_pin_memory: True
• dataloader_persistent_workers: False
• skip_memory_metrics: True
• use_legacy_prediction_loop: False
• push_to_hub: False
• resume_from_checkpoint: None
• hub_model_id: None
• hub_strategy: every_save
• hub_private_repo: None
• hub_always_push: False
• hub_revision: None
• gradient_checkpointing: True
• gradient_checkpointing_kwargs: None
• include_inputs_for_metrics: False
• include_for_metrics: []
• eval_do_concat_batches: True
• fp16_backend: auto
• push_to_hub_model_id: None
• push_to_hub_organization: None
• mp_parameters:
• auto_find_batch_size: False
• full_determinism: False
• torchdynamo: None
• ray_scope: last
• ddp_timeout: 1800
• torch_compile: False
• torch_compile_backend: None
• torch_compile_mode: None
• include_tokens_per_second: False
• include_num_input_tokens_seen: no
• neftune_noise_alpha: None
• optim_target_modules: None
• batch_eval_metrics: False
• eval_on_start: True
• use_liger_kernel: False
• liger_kernel_config: None
• eval_use_gather_object: False
• average_tokens_across_devices: True
• prompts: None
• batch_sampler: batch_sampler
• multi_dataset_batch_sampler: proportional
• router_mapping: {}
• learning_rate_mapping: {}

Training Logs

Click to expand

Environmental Impact

Carbon emissions were measured using CodeCarbon.

• Energy Consumed: 30.936 kWh
• Carbon Emitted: 6.350 kg of CO2
• Hours Used: 116.388 hours

Training Hardware

• On Cloud: No
• GPU Model: 1 x NVIDIA GeForce RTX 3090
• CPU Model: AMD Ryzen 7 3700X 8-Core Processor
• RAM Size: 62.70 GB

Framework Versions

• Python: 3.13.7
• Sentence Transformers: 5.1.2
• Transformers: 4.57.1
• PyTorch: 2.8.0+cu128
• Accelerate: 1.10.1
• Datasets: 4.3.0
• Tokenizers: 0.22.1

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",
}

Matryoshka2dLoss

@misc{li20242d,
    title={2D Matryoshka Sentence Embeddings},
    author={Xianming Li and Zongxi Li and Jing Li and Haoran Xie and Qing Li},
    year={2024},
    eprint={2402.14776},
    archivePrefix={arXiv},
    primaryClass={cs.CL}
}

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}
}

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},
}

TanimotoSentLoss

@online{cortes-2025-tanimotosentloss,
    title={TanimotoSentLoss: Tanimoto Loss for SMILES Embeddings},
    author={Emmanuel Cortes},
    year={2025},
    month={Jan},
    url={https://github.com/emapco/chem-mrl},
}

Model Card Authors

@eacortes

Model Card Contact

Manny Cortes (manny@derifyai.com)