metadata

base_model: microsoft/mpnet-base
datasets:
  - SwastikN/sxc_med_llm_chemical_gen
language:
  - en
library_name: sentence-transformers
license: apache-2.0
metrics:
  - cosine_accuracy
  - dot_accuracy
  - manhattan_accuracy
  - euclidean_accuracy
  - max_accuracy
pipeline_tag: sentence-similarity
tags:
  - sentence-transformers
  - sentence-similarity
  - feature-extraction
  - generated_from_trainer
  - dataset_size:117502
  - loss:MultipleNegativesRankingLoss
widget:
  - source_sentence: >-
      Help me make the molecule
      CC(=O)OC[C@H](OC(C)=O)C(=O)N1CCCC[C@H]1C1CCN(C(=O)c2cc3ccccc3n2C)CC1 with
      the same hydrogen bond donors. The output molecule should be similar to
      the input molecule. Please inform me of the number of hydrogen bond
      donor(s) of the optimized molecule.
    sentences:
      - >-
        Your requirements guided the optimization, resulting in the molecule
        "CC(=O)OC(CCl)C(Cc1cccs1)[C@H](OC(C)=O)C(=O)N1CCCC[C@H]1C1CCN(C(=O)c2cc3ccccc3n2C)CC1"
        with an approximate hydrogen bond donor(s) of 0.
      - >-
        Given a molecule expressed in SMILES string, help me optimize it
        according to my requirements.
      - >-
        Help me adapt a molecular structure denoted in SMILES string based on my
        preferences.
  - source_sentence: >-
      How can we modify the molecule
      CCC(CC)=C(CC)c1ccccc1OC(=O)OC(N=[N+]=[N-])c1ccccc1 to decrease its
      blood-brain barrier penetration (BBBP) value while keeping it similar to
      the input molecule? Please inform me of the BBBP value of the optimized
      molecule.
    sentences:
      - Describe a technology used for measuring people's emotional responses.
      - >-
        I've successfully optimized the molecule according to your needs,
        resulting in "CCOC(=O)c1ccccc1OC(=O)OC(N=[N+]=[N-])c1ccccc1" with an
        approximate BBBP value of 0.71.
      - >-
        Given a molecule expressed in SMILES string, help me optimize it
        according to my requirements.
  - source_sentence: >-
      How can we modify the molecule
      C/C(=C/C(=O)N1CC[C@H](CC(CCCCCC(CO)C(=O)O)NC(=O)OC(C)(C)C)[C@H]1c1cccnc1)C(=O)O
      to increase its blood-brain barrier penetration (BBBP) value while keeping
      it similar to the input molecule?
    sentences:
      - >-
        Given a molecule expressed in SMILES string, help me optimize it
        according to my requirements.
      - >-
        Aid me in refining a molecular structure written in SMILES notation
        based on my criteria.
      - >-
        Taking your requirements into account, I've optimized the molecule to
        "C/C(=C/C(=O)N1CC[C@H](CNC(=O)[C@H](CO)NC(=O)OC(C)(C)C)[C@H]1c1cccnc1)C(=O)O".
  - source_sentence: >-
      Support me in transforming the molecule [SMILES] by incorporating the same
      hydrogen bond acceptors and maintaining its resemblance to the original
      molecule.
    sentences:
      - >-
        Taking your requirements into account, I've optimized the molecule to
        "CCOc1cccc(C2c3c(oc4ccc(C)cc4c3=O)C(=O)N2CCN(CC)CC)c1".
      - >-
        Help me adapt a molecular structure denoted in SMILES string based on my
        preferences.
      - >-
        Help me adapt a molecular structure denoted in SMILES string based on my
        preferences.
  - source_sentence: >-
      With a molecule represented by the SMILES string
      CNNNCC(=O)N[C@H](C)C[C@@H](C)NCc1ccc2c(c1)CCC2, propose adjustments that
      can increase its logP value while keeping the output molecule structurally
      related to the input molecule.
    sentences:
      - >-
        Aid me in refining a molecular structure written in SMILES notation
        based on my criteria.
      - >-
        Given a molecule expressed in SMILES string, help me optimize it
        according to my requirements.
      - >-
        In line with your criteria, I've optimized the molecule and present it
        as "C[C@H](C[C@@H](C)NC(=O)COC(C)(C)C)NCc1ccc2c(c1)CCC2".
model-index:
  - name: MPNet base trained on AllNLI triplets
    results:
      - task:
          type: triplet
          name: Triplet
        dataset:
          name: all nli dev
          type: all-nli-dev
        metrics:
          - type: cosine_accuracy
            value: 0.6562222222222223
            name: Cosine Accuracy
          - type: dot_accuracy
            value: 0.5342222222222223
            name: Dot Accuracy
          - type: manhattan_accuracy
            value: 0.7075555555555556
            name: Manhattan Accuracy
          - type: euclidean_accuracy
            value: 0.6584444444444445
            name: Euclidean Accuracy
          - type: max_accuracy
            value: 0.7075555555555556
            name: Max Accuracy
          - type: cosine_accuracy
            value: 0.9804444444444445
            name: Cosine Accuracy
          - type: dot_accuracy
            value: 0.01888888888888889
            name: Dot Accuracy
          - type: manhattan_accuracy
            value: 0.9811111111111112
            name: Manhattan Accuracy
          - type: euclidean_accuracy
            value: 0.9802222222222222
            name: Euclidean Accuracy
          - type: max_accuracy
            value: 0.9811111111111112
            name: Max Accuracy

MPNet base trained on AllNLI triplets

This is a sentence-transformers model finetuned from microsoft/mpnet-base on the sxc_med_llm_chemical_gen dataset. It maps sentences & paragraphs to a 768-dimensional dense vector space and can be used for semantic textual similarity, semantic search, paraphrase mining, text classification, clustering, and more.

Model Details

Model Description

Model Type: Sentence Transformer
Base model: microsoft/mpnet-base
Maximum Sequence Length: 512 tokens
Output Dimensionality: 768 tokens
Similarity Function: Cosine Similarity
Training Dataset:
- sxc_med_llm_chemical_gen
Language: en
License: apache-2.0

Model Sources

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

Full Model Architecture

SentenceTransformer(
  (0): Transformer({'max_seq_length': 512, 'do_lower_case': False}) with Transformer model: MPNetModel 
  (1): Pooling({'word_embedding_dimension': 768, '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})
)

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("Saideepthi55/sentencetransformer-ft")
# Run inference
sentences = [
    'With a molecule represented by the SMILES string CNNNCC(=O)N[C@H](C)C[C@@H](C)NCc1ccc2c(c1)CCC2, propose adjustments that can increase its logP value while keeping the output molecule structurally related to the input molecule.',
    'Given a molecule expressed in SMILES string, help me optimize it according to my requirements.',
    'In line with your criteria, I\'ve optimized the molecule and present it as "C[C@H](C[C@@H](C)NC(=O)COC(C)(C)C)NCc1ccc2c(c1)CCC2".',
]
embeddings = model.encode(sentences)
print(embeddings.shape)
# [3, 768]

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

Evaluation

Metrics

Triplet

Dataset: all-nli-dev
Evaluated with TripletEvaluator

Metric	Value
cosine_accuracy	0.6562
dot_accuracy	0.5342
manhattan_accuracy	0.7076
euclidean_accuracy	0.6584
max_accuracy	0.7076

Triplet

Dataset: all-nli-dev
Evaluated with TripletEvaluator

Metric	Value
cosine_accuracy	0.9804
dot_accuracy	0.0189
manhattan_accuracy	0.9811
euclidean_accuracy	0.9802
max_accuracy	0.9811

Training Details

Training Dataset

sxc_med_llm_chemical_gen

Dataset: sxc_med_llm_chemical_gen at ae28874
Size: 117,502 training samples
Columns: input, instruction, and output

Approximate statistics based on the first 1000 samples:

	input	instruction	output
type	string	string	string
details	min: 2 tokens mean: 50.03 tokens max: 156 tokens	min: 6 tokens mean: 16.19 tokens max: 69 tokens	min: 3 tokens mean: 67.94 tokens max: 321 tokens

Samples:

input	instruction	output
	`List the countries of the Middle East`	`Egypt, Israel, Iran, Iraq, Jordan, Kuwait, Lebanon, Qatar, Saudi Arabia, Syria, Turkey, The United Arab Emirates, Oman, and Yemen.`
`Support me in refining the molecule CCC(C)=C(CC)c1ccc(Nc2nc3nonc3nc2N2CCC(Oc3ccc(Cl)cc3)CC2)cc1 to have a more drug-like nature, ensuring it remains close to the input molecule. Please inform me of the QED value of the optimized molecule.`	`Help me adapt a molecular structure denoted in SMILES string based on my preferences.`	`I've optimized the molecule according to your requirements, yielding the structure "COC(=O)c1ccc(Nc2nc3nonc3nc2N2CCC(Oc3ccc(Cl)cc3)CC2)cc1" and an estimated QED value of 0.4.`
	`Create a social media strategy to attract followers on Twitter.`	The social media strategy for Twitter should involve Tweeting on a regular basis, with content that is engaging and relevant to the target audience. This can include retweetable quotes, links to videos and articles, sharing photos/images, polls, and so on. It is also beneficial for the information to include hashtags relevant to the content and any relevant topics of the industry. It is also important to leverage Twitter influencers and accounts, as well as engaging with potential followers regularly and DM them when necessary. Additionally, running campaigns, contests and giveaways to gain more followers should also be part of the strategy. Finally, hashtag research should be conducted to increase the reach of the tweets.

Loss: MultipleNegativesRankingLoss with these parameters:

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

Evaluation Dataset

sxc_med_llm_chemical_gen

Dataset: sxc_med_llm_chemical_gen at ae28874
Size: 4,500 evaluation samples
Columns: input, instruction, and output

Approximate statistics based on the first 1000 samples:

	input	instruction	output
type	string	string	string
details	min: 2 tokens mean: 68.49 tokens max: 165 tokens	min: 6 tokens mean: 16.89 tokens max: 36 tokens	min: 3 tokens mean: 69.16 tokens max: 205 tokens

Samples:

input	instruction	output
`With a molecule represented by the SMILES string COc1ccc(CN(C(=O)Nc2ccc(C@Hc3ccccc3)c(Cl)c2)C2CC2)cc1, propose adjustments that can decrease its blood-brain barrier permeability (BBBP) value by at least 0.1 compared to the pre-optimized value while keeping the output molecule structurally related to the input molecule. Please inform me of the BBBP value of the optimized molecule.`	`Help me adapt a molecular structure denoted in SMILES string based on my preferences.`	`I've optimized the molecule according to your requirements, yielding the structure "CC(C)c1cccc(Oc2ccc(CN(C(=O)Nc3ccc(C@Hc4ccccc4)c(Cl)c3)C3CC3)cc2)c1" and an estimated BBBP value of 0.34.`
`Which modifications can be applied to the molecule Cc1nn(-c2cc3cc(F)ccc3s2)c2ncc(C(=O)N3CCCC3c3nnc4n3CCCCC4)cc12 to increase its water solubility value while ensuring it remains similar to the original molecule?`	`Aid me in refining a molecular structure written in SMILES notation based on my criteria.`	`I've optimized the molecule according to your requirements, resulting in the following structure: "Cc1nn(C)c2ncc(C(=O)N3CCCC3c3nnc4n3CCCCC4)cc12".`
	`Describe the benefits of having a mentor.`	`Having a mentor can provide tremendous benefits, including guidance, advice, support and connection to a network of contacts. Mentors can offer valuable feedback, help to create goals and strategies, and provide the accountability necessary to carry out the desired goals. They can also provide a fresh perspective which can help to create new ideas and solutions.`

Loss: MultipleNegativesRankingLoss with these parameters:

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

Training Hyperparameters

Non-Default Hyperparameters

eval_strategy: steps
learning_rate: 2e-05
num_train_epochs: 1
warmup_ratio: 0.1
fp16: 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: 8
per_device_eval_batch_size: 8
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: 2e-05
weight_decay: 0.0
adam_beta1: 0.9
adam_beta2: 0.999
adam_epsilon: 1e-08
max_grad_norm: 1.0
num_train_epochs: 1
max_steps: -1
lr_scheduler_type: linear
lr_scheduler_kwargs: {}
warmup_ratio: 0.1
warmup_steps: 0
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: None
jit_mode_eval: False
use_ipex: False
bf16: False
fp16: True
fp16_opt_level: O1
half_precision_backend: auto
bf16_full_eval: False
fp16_full_eval: False
tf32: None
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}
deepspeed: None
label_smoothing_factor: 0.0
optim: adamw_torch
optim_args: None
adafactor: False
group_by_length: False
length_column_name: length
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: False
hub_always_push: False
gradient_checkpointing: False
gradient_checkpointing_kwargs: None
include_inputs_for_metrics: False
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
dispatch_batches: None
split_batches: None
include_tokens_per_second: False
include_num_input_tokens_seen: False
neftune_noise_alpha: None
optim_target_modules: None
batch_eval_metrics: False
eval_on_start: False
eval_use_gather_object: False
batch_sampler: batch_sampler
multi_dataset_batch_sampler: proportional

Training Logs

Epoch	Step	Training Loss	loss	all-nli-dev_max_accuracy
0	0	-	-	0.7076
0.0174	64	-	-	0.7156
0.0068	100	2.7336	2.6486	0.7524
0.0136	200	2.4965	1.9213	0.8162
0.0204	300	1.9042	1.7761	0.822
0.0272	400	1.6856	1.7172	0.8371
0.0340	500	1.6117	1.6916	0.8507
0.0408	600	1.5673	1.6809	0.8976
0.0477	700	1.5984	1.7052	0.9329
0.0545	800	1.5828	1.6841	0.9391
0.0613	900	1.5375	1.6534	0.9267
0.0681	1000	1.5561	1.6619	0.9509
0.0749	1100	1.4911	1.6538	0.9556
0.0817	1200	1.5075	1.6498	0.966
0.0885	1300	1.4722	1.6468	0.946
0.0953	1400	1.4806	1.6981	0.9631
0.1021	1500	1.4788	1.6335	0.9662
0.1089	1600	1.4668	1.6668	0.9731
0.1157	1700	1.4383	1.6473	0.9711
0.1225	1800	1.4549	1.6462	0.9713
0.1294	1900	1.4394	1.6184	0.9718
0.1362	2000	1.3861	1.6156	0.9676
0.1430	2100	1.4111	1.6045	0.9711
0.1498	2200	1.4286	1.6056	0.9782
0.1566	2300	1.4669	1.6174	0.9764
0.1634	2400	1.3761	1.6182	0.9776
0.1702	2500	1.4119	1.6150	0.9738
0.1770	2600	1.3625	1.5984	0.9776
0.1838	2700	1.3726	1.6092	0.9807
0.1906	2800	1.3265	1.6059	0.9789
0.1974	2900	1.3925	1.6004	0.978
0.2042	3000	1.3524	1.5964	0.9773
0.2111	3100	1.342	1.6213	0.9787
0.2179	3200	1.3478	1.6016	0.9822
0.2247	3300	1.3888	1.6038	0.9793
0.2315	3400	1.3328	1.5977	0.9813
0.2383	3500	1.372	1.6114	0.9824
0.2451	3600	1.3046	1.6082	0.9824
0.2519	3700	1.3857	1.5922	0.9824
0.2587	3800	1.3236	1.6127	0.9809
0.2655	3900	1.2929	1.5935	0.9824
0.2723	4000	1.3889	1.6047	0.9831
0.2791	4100	1.3509	1.6030	0.9844
0.2859	4200	1.3455	1.6099	0.9824
0.2928	4300	1.337	1.5939	0.984
0.2996	4400	1.3302	1.6057	0.9827
0.3064	4500	1.3377	1.6254	0.9833
0.3132	4600	1.3221	1.6020	0.9849
0.3200	4700	1.3209	1.6146	0.9824
0.3268	4800	1.354	1.6022	0.9824
0.3336	4900	1.3213	1.6136	0.9822
0.3404	5000	1.3484	1.5920	0.9807
0.3472	5100	1.3412	1.6106	0.978
0.3540	5200	1.3532	1.6001	0.9784
0.3608	5300	1.2984	1.6192	0.9762
0.3676	5400	1.3621	1.5850	0.98
0.3745	5500	1.2839	1.6158	0.9807
0.3813	5600	1.3664	1.6030	0.9831
0.3881	5700	1.327	1.6168	0.9822
0.3949	5800	1.3123	1.6040	0.982
0.4017	5900	1.3019	1.6092	0.9824
0.4085	6000	1.3908	1.5935	0.9829
0.4153	6100	1.3136	1.5916	0.9791
0.4221	6200	1.32	1.6091	0.9807
0.4289	6300	1.3018	1.6052	0.9827
0.4357	6400	1.3144	1.6083	0.9816
0.4425	6500	1.2865	1.6015	0.9829
0.4493	6600	1.2946	1.5882	0.9818
0.4562	6700	1.3245	1.5949	0.9824
0.4630	6800	1.3278	1.6081	0.9831
0.4698	6900	1.2842	1.6086	0.9836
0.4766	7000	1.3231	1.6170	0.9811

Framework Versions

Python: 3.10.12
Sentence Transformers: 3.1.0
Transformers: 4.44.2
PyTorch: 2.4.0+cu121
Accelerate: 0.34.2
Datasets: 3.0.0
Tokenizers: 0.19.1

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",
}

MultipleNegativesRankingLoss

@misc{henderson2017efficient,
    title={Efficient Natural Language Response Suggestion for Smart Reply},
    author={Matthew Henderson and Rami Al-Rfou and Brian Strope and Yun-hsuan Sung and Laszlo Lukacs and Ruiqi Guo and Sanjiv Kumar and Balint Miklos and Ray Kurzweil},
    year={2017},
    eprint={1705.00652},
    archivePrefix={arXiv},
    primaryClass={cs.CL}
}