Derify/ChemRanker-alpha-sim

This Cross Encoder is finetuned from Derify/ModChemBERT-IR-BASE using hard-negative triplets derived from Derify/pubchem_10m_genmol_similarity. Positive SMILES pairs are first filtered by quality and similarity constraints, then reduced to one strongest positive target per anchor molecule to create a high-signal training set for reranking. The model computes relevance scores for pairs of SMILES strings, enabling SMILES reranking and molecular semantic search.

For this variant, the positive selection objective is pure similarity ranking where each anchor keeps the highest-similarity candidate after filtering, rather than using a QED+similarity composite score. The quality stage uses strict inequality filtering (QED > 0.85, similarity > 0.5, with similarity also bounded below 1.0), and then keeps the top-scoring pair per anchor molecule.

Hard negatives are mined with Sentence Transformers using Derify/ChemMRL-beta as the teacher model and a TopK-PercPos-style margin setting based on NV-Retriever, with relative_margin=0.05 and max_negative_score_threshold = pos_score * percentage_margin. Training uses triplet-format samples with 5 mined negatives per anchor-positive pair and optimizes a multiple-negatives ranking objective, while reranking evaluation uses n-tuple samples with 30 mined negatives per query.

Model Details

Model Description

• Model Type: Cross Encoder
• Base model: Derify/ModChemBERT-IR-BASE
• Maximum Sequence Length: 512 tokens
• Number of Output Labels: 1 label
• Training Dataset:
  • Derify/pubchem_10m_genmol_similarity Mined Hard Negatives
• License: apache-2.0

Model Sources

• Documentation: Sentence Transformers Documentation
• Documentation: Cross Encoder Documentation
• Repository: Sentence Transformers on GitHub
• Hugging Face: Cross Encoders on Hugging Face

Usage

Direct Usage (Sentence Transformers)

First install the Transformers and Sentence Transformers libraries:

pip install -U "transformers>=4.57.1,<5.0.0"
pip install -U sentence-transformers

Then you can load this model and run inference.

from sentence_transformers import CrossEncoder

# Download from the 🤗 Hub
model = CrossEncoder("Derify/ChemRanker-alpha-sim")
# Get scores for pairs of texts
pairs = [
    ['c1snnc1C[NH2+]Cc1cc2c(s1)CCC2', 'c1snnc1CCC[NH2+]Cc1cc2c(s1)CCC2'],
    ['c1sc2c(c1-c1nc(C3CCOC3)no1)CCCC2', 'O=CCc1noc(-c2csc3c2CCCC3)n1'],
    ['c1sc(C[NH2+]C2CC2)nc1C[NH+]1CCN2CCCC2C1', 'FC(F)[NH2+]Cc1nc(C[NH+]2CCN3CCCC3C2)cs1'],
    ['c1sc(CC[NH+]2CCOCC2)nc1C[NH2+]C1CC1', 'CCc1nc(C[NH2+]C2CC2)cs1'],
    ['c1sc(CC2CCC[NH2+]2)nc1C1CCCO1', 'c1sc(CC2CCC[NH2+]2)nc1C1CCCC1'],
]
scores = model.predict(pairs)
print(scores.shape)
# (5,)

# Or rank different texts based on similarity to a single text
ranks = model.rank(
    'c1snnc1C[NH2+]Cc1cc2c(s1)CCC2',
    [
        'c1snnc1CCC[NH2+]Cc1cc2c(s1)CCC2',
        'O=CCc1noc(-c2csc3c2CCCC3)n1',
        'FC(F)[NH2+]Cc1nc(C[NH+]2CCN3CCCC3C2)cs1',
        'CCc1nc(C[NH2+]C2CC2)cs1',
        'c1sc(CC2CCC[NH2+]2)nc1C1CCCC1',
    ]
)
# [{'corpus_id': ..., 'score': ...}, {'corpus_id': ..., 'score': ...}, ...]

Evaluation

Metrics

Cross Encoder Reranking

• Evaluated with CrossEncoderRerankingEvaluator with these parameters:

  {
      "at_k": 10
  }
  

Metric	Value
map	0.4323
mrr@10	0.6975
ndcg@10	0.7034

Training Details

Training Dataset

GenMol Similarity Hard Negatives

• Dataset: GenMol Similarity Hard Negatives
• Size: 3,269,544 training samples
• Columns: smiles_a, smiles_b, and negative
• Approximate statistics based on the first 1000 samples:

  	smiles_a	smiles_b	negative
  type	string	string	string
  details	min: 19 characters mean: 33.64 characters max: 65 characters	min: 20 characters mean: 34.16 characters max: 54 characters	min: 19 characters mean: 33.28 characters max: 57 characters

• Samples:

  smiles_a	smiles_b	negative
  c1sc2cc3c(cc2c1CC[NH2+]C1CC1)OCCO3	FC(F)(F)[NH2+]CCc1csc2cc3c(cc12)OCCO3	[NH3+]CCCc1cc2c(cc1C1CC1)OCO2
  c1sc2cc3c(cc2c1CC[NH2+]C1CC1)OCCO3	FC(F)(F)[NH2+]CCc1csc2cc3c(cc12)OCCO3	COc1cc2c(cc1C[NH2+]C1CCC1)OCO2
  c1sc2cc3c(cc2c1CC[NH2+]C1CC1)OCCO3	FC(F)(F)[NH2+]CCc1csc2cc3c(cc12)OCCO3	O=c1[nH]c2cc3c(cc2cc1CNC1CCCCC1)OCCO3

• Loss: MultipleNegativesRankingLoss with these parameters:

  {
      "scale": 10.0,
      "num_negatives": 4,
      "activation_fn": "torch.nn.modules.activation.Sigmoid"
  }
  

Evaluation Dataset

GenMol Similarity Hard Negatives

• Dataset: GenMol Similarity Hard Negatives
• Size: 165,968 evaluation samples
• Columns: smiles_a, smiles_b, negative_1, negative_2, negative_3, negative_4, negative_5, negative_6, negative_7, negative_8, negative_9, negative_10, negative_11, negative_12, negative_13, negative_14, negative_15, negative_16, negative_17, negative_18, negative_19, negative_20, negative_21, negative_22, negative_23, negative_24, negative_25, negative_26, negative_27, negative_28, negative_29, and negative_30
• Approximate statistics based on the first 1000 samples:

  	smiles_a	smiles_b	negative_1	negative_2	negative_3	negative_4	negative_5	negative_6	negative_7	negative_8	negative_9	negative_10	negative_11	negative_12	negative_13	negative_14	negative_15	negative_16	negative_17	negative_18	negative_19	negative_20	negative_21	negative_22	negative_23	negative_24	negative_25	negative_26	negative_27	negative_28	negative_29	negative_30
  type	string	string	string	string	string	string	string	string	string	string	string	string	string	string	string	string	string	string	string	string	string	string	string	string	string	string	string	string	string	string	string	string
  details	min: 17 characters mean: 37.57 characters max: 96 characters	min: 14 characters mean: 34.46 characters max: 70 characters	min: 16 characters mean: 35.94 characters max: 77 characters	min: 12 characters mean: 35.1 characters max: 77 characters	min: 14 characters mean: 35.09 characters max: 81 characters	min: 17 characters mean: 35.38 characters max: 74 characters	min: 17 characters mean: 35.17 characters max: 70 characters	min: 14 characters mean: 35.25 characters max: 84 characters	min: 16 characters mean: 35.2 characters max: 77 characters	min: 13 characters mean: 35.05 characters max: 80 characters	min: 11 characters mean: 35.25 characters max: 90 characters	min: 11 characters mean: 35.23 characters max: 74 characters	min: 12 characters mean: 34.88 characters max: 60 characters	min: 14 characters mean: 35.42 characters max: 66 characters	min: 13 characters mean: 35.36 characters max: 69 characters	min: 13 characters mean: 34.81 characters max: 77 characters	min: 10 characters mean: 35.12 characters max: 77 characters	min: 17 characters mean: 35.05 characters max: 69 characters	min: 14 characters mean: 35.47 characters max: 72 characters	min: 14 characters mean: 35.12 characters max: 65 characters	min: 18 characters mean: 35.44 characters max: 72 characters	min: 14 characters mean: 35.0 characters max: 64 characters	min: 18 characters mean: 35.79 characters max: 81 characters	min: 17 characters mean: 35.43 characters max: 67 characters	min: 14 characters mean: 35.76 characters max: 68 characters	min: 14 characters mean: 35.29 characters max: 62 characters	min: 17 characters mean: 35.42 characters max: 66 characters	min: 16 characters mean: 35.31 characters max: 83 characters	min: 18 characters mean: 35.64 characters max: 77 characters	min: 18 characters mean: 35.47 characters max: 77 characters	min: 11 characters mean: 35.23 characters max: 65 characters	min: 16 characters mean: 35.26 characters max: 77 characters

• Samples:

  smiles_a	smiles_b	negative_1	negative_2	negative_3	negative_4	negative_5	negative_6	negative_7	negative_8	negative_9	negative_10	negative_11	negative_12	negative_13	negative_14	negative_15	negative_16	negative_17	negative_18	negative_19	negative_20	negative_21	negative_22	negative_23	negative_24	negative_25	negative_26	negative_27	negative_28	negative_29	negative_30
  c1snnc1C[NH2+]Cc1cc2c(s1)CCC2	c1snnc1CCC[NH2+]Cc1cc2c(s1)CCC2	c1snnc1CCC[NH2+]Cc1cc2c(s1)CCC2	Cn1cc(C[NH2+]Cc2cc3c(s2)CCC3)nn1	Cn1cc(CC[NH2+]Cc2cc3c(s2)CCC3)nn1	Cc1cc(C[NH2+]Cc2csnn2)sc1C	NC(=O)c1csc(C[NH2+]Cc2cc3c(s2)CCC3)c1	Cc1cc(CC[NH2+]Cc2csnn2)sc1C	Ic1ccc(C[NH2+]Cc2cc3c(s2)CCC3)o1	Cc1cc(C[NH2+]CCCCc2cc3c(s2)CCC3)c(C)s1	c1ccc(C[NH2+]Cc2cc3c(s2)CCC3)cc1	c1ncc(C[NH2+]Cc2csnn2)s1	FC(F)c1csc(C[NH2+]Cc2cc3c(s2)CCC3)c1	c1c(C[NH2+]CC2CC2)sc2c1CSCC2	N#Cc1cc(F)cc(C[NH2+]Cc2cc3c(s2)CCC3)c1	c1cc(C[NH2+]Cc2nc3c(s2)CCC3)no1	CCc1ccc(C[NH2+]Cc2csnn2)s1	NCc1csc(NCc2cc3c(s2)CCC3)n1	CNH+Cc1nnc(-c2cc3c(s2)CCCC3)o1	Fc1cc(C[NH2+]Cc2cc3c(s2)CCC3)ccc1Br	FC(F)(F)C[NH2+]Cc1cc2c(s1)CCSC2	c1cc(C[NH2+]Cc2cc3c(s2)CCC3)c[nH]1	Cc1cc(C)c(CC[NH2+]Cc2cc3c(s2)CCC3)c(C)c1	Oc1ccc(C[NH2+]Cc2cc3c(s2)CCC3)cc1Br	O=C([O-])c1ccc(CC[NH2+]Cc2cc3c(s2)CCC3)s1	c1c(C[NH2+]CC2CCCC2)sc2c1CCC2	O=C([O-])c1ccc(C[NH2+]Cc2cc3c(s2)CCC3)s1	COc1cc(C)cc(C[NH2+]Cc2cc3c(s2)CCC3)c1	PSc1ccc(C[NH2+]Cc2csnn2)s1	CCc1cnc(C[NH2+]Cc2csnn2)s1	Clc1cc(C[NH2+]Cc2cc3c(s2)CCC3)ccc1Br	c1c(C[NH2+]CC2CC2)sc2c1CCCCC2
  c1sc2c(c1-c1nc(C3CCOC3)no1)CCCC2	O=CCc1noc(-c2csc3c2CCCC3)n1	Nc1sc2c(c1-c1nc(C3CCOC3)no1)CCCC2	Nc1sc2c(c1-c1nc(C3CCC3)no1)CCCC2	c1c(-c2nc(C3CCCNC3)no2)sc2c1CCCCCC2	Nc1sccc1-c1nc(C2CCCOC2)no1	Nc1sc2c(c1-c1nc(C3CCCO3)no1)CCCC2	Cc1csc(-c2nc(C3CCOCC3)no2)c1N	Cc1oc2c(c1-c1nc(C3CCOC3)no1)C(=O)CCC2	c1c(-c2nc(C3C[NH2+]CCO3)no2)sc2c1CCCCC2	O=C([O-])Nc1sc2c(c1-c1nc(C3CC3)no1)CCCC2	c1cc2c(s1)CCCC2c1nc(C2CC2)no1	CC(=O)N1CCCC(c2noc(-c3cc4c(s3)CCCCCC4)n2)C1	Cc1cc(-c2nc([C@@H]3CCOC3)no2)c(N)s1	c1cc2c(nc1-c1noc(C3CCCOC3)n1)CCCC2	Nc1sccc1-c1nc(C2CCCC2)no1	c1cc2c(nc1-c1noc(C3CCOCC3)n1)CCCC2	[NH3+]C(c1noc(-c2cc3c(s2)CCCC3)n1)C1CC1	c1cc2c(c(-c3nc(C4CCOCC4)no3)c1)CCCN2	c1c(-c2nc(C3CC3)no2)nn2c1CCCC2	CN1CC(c2noc(-c3cc4c(s3)CCCC4)n2)CC1=O	Oc1c(-c2nc(C3CCC(F)(F)C3)no2)ccc2c1CCCC2	O=CCc1noc(-c2csc3c2CCCC3)n1	Cc1cc(=O)c(-c2noc(C3CCCOC3)n2)c2n1CCC2	O=C([O-])CNc1sc2c(c1-c1nc(C3CC3)no1)CCCC2	c1cc(-c2noc(C3CCCOC3)n2)cs1	Cn1nc(-c2nc(C3CCCO3)no2)c2c1CCCC2	O=C(Nc1sc2c(c1-c1nc(C3CC3)no1)COCC2)C1=CCCCC1	Cc1cscc1-c1noc(C2CCOCC2)n1	CC1(C)CCCc2sc(N)c(-c3nc(C4CC4)no3)c21	Clc1cc2c(c(-c3nc(C4CCOC4)no3)c1)OCC2	Nc1sc2c(c1-c1nnc(C3CC3)o1)CCCC2
  c1sc(C[NH2+]C2CC2)nc1C[NH+]1CCN2CCCC2C1	FC(F)[NH2+]Cc1nc(C[NH+]2CCN3CCCC3C2)cs1	FC(F)[NH2+]Cc1nc(C[NH+]2CCN3CCCC3C2)cs1	CC(C)[NH2+]Cc1nc(C[NH+]2CCC3CCCCC3C2)cs1	CN1C2CCC1CNH+CC2	Nc1nc(CC[NH+]2CCCN3CCCC3C2)cs1	NCc1nc(C[NH+]2CCCC3CCCCC32)cs1	CC1CNH+CCN1C	Oc1csc(CN2CCCC3C[NH2+]CC32)n1	CCc1nc(C[NH+]2CCCC3CCCCC32)cs1	C[NH2+]Cc1csc(N2CC[NH+]3CCCC3C2)n1	[NH3+]Cc1nc(C[NH+]2CCC3CCCCC32)cs1	CC1CN2CCCCC2C[NH+]1Cc1csc(CC[NH3+])n1	CCCc1nc(CN2CCCC2C2CCC[NH2+]2)cs1	ClCCc1nc(CN2CCCC2C2CCC[NH2+]2)cs1	c1cc(C[NH2+]C2CC2)c(C[NH+]2CCN3CCCCC3C2)o1	O=C(Cc1nc(CCl)cs1)N1CCC[NH+]2CCCC2C1	CC[NH2+]Cc1csc(N2CCC3C(CCC[NH+]3C)C2)n1	c1sc(C[NH2+]C2CC2)nc1C[NH+]1CCCCC1	[NH3+]Cc1nc(C[NH+]2CCCC2C2CCCC2)cs1	Cc1csc(C[NH+]2CCC3C[NH2+]CC3C2)n1	c1cc(C[NH+]2CCCN3CCCC3C2)nc(C2CC2)n1	Cc1ccsc1C[NH2+]CCN1CCN2CCCC2C1	c1sc(C[NH2+]C2CCCC2)nc1C[NH+]1CCCCC1	Brc1csc(C[NH2+]CCN2CCN3CCCCC3C2)c1	Cc1nc(CCC[NH2+]C2CCN3CCCCC23)cs1	CCOC(=O)c1nc(CN2CC3CCC[NH2+]C3C2)cs1	CCCC(=O)c1nc(CN2CC3CCC[NH2+]C3C2)cs1	CC(C)(C)c1csc(CN2CCC[NH2+]C(C3CC3)C2)n1	COCc1nc(CN2CCC([NH3+])C2)cs1	CCC[NH2+]Cc1nc(C[NH+]2CC3CCC2C3)cs1	CCC1CN2CCCC2C[NH+]1CCc1csc(C)n1

• Loss: MultipleNegativesRankingLoss with these parameters:

  {
      "scale": 10.0,
      "num_negatives": 4,
      "activation_fn": "torch.nn.modules.activation.Sigmoid"
  }
  

Training Hyperparameters

Non-Default Hyperparameters

• eval_strategy: epoch
• per_device_train_batch_size: 256
• per_device_eval_batch_size: 256
• torch_empty_cache_steps: 1000
• learning_rate: 3e-05
• weight_decay: 1e-05
• max_grad_norm: None
• lr_scheduler_type: warmup_stable_decay
• lr_scheduler_kwargs: {'num_decay_steps': 6385, 'warmup_type': 'linear', 'decay_type': '1-sqrt'}
• warmup_steps: 6385
• seed: 12
• data_seed: 24681357
• bf16: True
• bf16_full_eval: True
• tf32: True
• dataloader_num_workers: 8
• dataloader_prefetch_factor: 2
• load_best_model_at_end: True
• optim: stable_adamw
• optim_args: decouple_lr=True,max_lr=3e-05
• dataloader_persistent_workers: True
• resume_from_checkpoint: False
• gradient_checkpointing: True
• torch_compile: True
• torch_compile_backend: inductor
• torch_compile_mode: max-autotune
• eval_on_start: True
• batch_sampler: no_duplicates

All Hyperparameters

Click to expand

• overwrite_output_dir: False
• do_predict: False
• eval_strategy: epoch
• prediction_loss_only: True
• per_device_train_batch_size: 256
• per_device_eval_batch_size: 256
• per_gpu_train_batch_size: None
• per_gpu_eval_batch_size: None
• gradient_accumulation_steps: 1
• eval_accumulation_steps: None
• torch_empty_cache_steps: 1000
• learning_rate: 3e-05
• 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': 6385, 'warmup_type': 'linear', 'decay_type': '1-sqrt'}
• warmup_ratio: 0.0
• warmup_steps: 6385
• 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: 12
• data_seed: 24681357
• 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: True
• dataloader_num_workers: 8
• dataloader_prefetch_factor: 2
• past_index: -1
• disable_tqdm: False
• remove_unused_columns: True
• label_names: None
• load_best_model_at_end: True
• 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=3e-05
• 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: True
• skip_memory_metrics: True
• use_legacy_prediction_loop: False
• push_to_hub: False
• resume_from_checkpoint: False
• 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: True
• torch_compile_backend: inductor
• torch_compile_mode: max-autotune
• 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: no_duplicates
• multi_dataset_batch_sampler: proportional
• router_mapping: {}
• learning_rate_mapping: {}

Training Logs

Epoch	Step	Training Loss	Validation Loss	ndcg@10
1.0963	7000	0.0046	-	-
1.2529	8000	0.0043	-	-
1.4096	9000	0.0038	-	-
1.5662	10000	0.0035	-	-
1.7228	11000	0.0033	-	-
1.8794	12000	0.0031	-	-
2.0	12770	-	1.5814	0.6986
2.0360	13000	0.003	-	-
2.1926	14000	0.0027	-	-
2.3493	15000	0.0025	-	-
2.5059	16000	0.0025	-	-
2.6625	17000	0.0024	-	-
2.8191	18000	0.0024	-	-
2.9757	19000	0.0024	-	-
3.0	19155	-	1.5688	0.7034

• The bold row denotes the saved checkpoint.

Environmental Impact

Carbon emissions were measured using CodeCarbon.

• Energy Consumed: 12.236 kWh
• Carbon Emitted: 2.512 kg of CO2
• Hours Used: 19.958 hours

Training Hardware

• On Cloud: No
• GPU Model: 2 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: 2.9.0+cu128
• Accelerate: 1.11.0
• Datasets: 4.4.1
• 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",
}

NV-Retriever

@misc{moreira2025nvretrieverimprovingtextembedding,
      title={NV-Retriever: Improving text embedding models with effective hard-negative mining}, 
      author={Gabriel de Souza P. Moreira and Radek Osmulski and Mengyao Xu and Ronay Ak and Benedikt Schifferer and Even Oldridge},
      year={2025},
      eprint={2407.15831},
      archivePrefix={arXiv},
      primaryClass={cs.IR},
      url={https://arxiv.org/abs/2407.15831}, 
}