Aditya109/drug-embed-india-v2

SentenceTransformer based on FremyCompany/BioLORD-2023

India Medicine Brand Research Model

This experimental model was fine-tuned to place Indian medicine brand names with the same recorded active-ingredient composition closer together. It is intended for research, catalog-search experiments, and model evaluation only.

It must not be used for diagnosis, prescribing, dosage decisions, interaction checks, or automatic medicine substitution. Training labels came from an automatically processed public A-Z medicines dataset and were not independently reviewed by pharmacists. Verify the source dataset's license before commercial use.

Fresh held-out evaluation against the BioLORD base model:

• nearest-neighbour composition accuracy: 0.2394 -> 0.4859
• validation triplet accuracy: 0.3322 -> 0.5242
• look-alike triplet accuracy: 0.1685 -> 0.3875
• unseen-composition triplet accuracy: 0.2690 -> 0.4342

This is a sentence-transformers model finetuned from FremyCompany/BioLORD-2023 on the pairs and triplets datasets. It maps sentences & paragraphs to a 768-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
• Base model: FremyCompany/BioLORD-2023
• Maximum Sequence Length: 32 tokens
• Output Dimensionality: 768 dimensions
• Similarity Function: Cosine Similarity
• Supported Modality: Text
• Training Datasets:
  • pairs
  • triplets

Model Sources

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

Full Model Architecture

SentenceTransformer(
  (0): Transformer({'transformer_task': 'feature-extraction', 'modality_config': {'text': {'method': 'forward', 'method_output_name': 'last_hidden_state'}}, 'module_output_name': 'token_embeddings', 'architecture': 'MPNetModel'})
  (1): Pooling({'embedding_dimension': 768, 'pooling_mode': 'mean', '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("sentence_transformers_model_id")
# Run inference
sentences = [
    'Docrox 90mg',
    'Etoricoxib',
    'Letroma 2.5',
]
embeddings = model.encode(sentences)
print(embeddings.shape)
# [3, 768]

# Get the similarity scores for the embeddings
similarities = model.similarity(embeddings, embeddings)
print(similarities)
# tensor([[1.0000, 0.8378, 0.6347],
#         [0.8378, 1.0000, 0.6442],
#         [0.6347, 0.6442, 1.0000]])

Evaluation

Metrics

Composition

• Dataset: comp
• Evaluated with main.CompositionEvaluator

Metric	Value
same	0.7836
diff	0.6976
margin	0.086
margin_norm	1.6511
nn_accuracy	0.4789

Triplet

• Datasets: val, lookalike and unseen_comp
• Evaluated with TripletEvaluator

Metric	val	lookalike	unseen_comp
cosine_accuracy	0.533	0.397	0.444

Training Details

Training Datasets

pairs

• Dataset: pairs
• Size: 371,947 training samples
• Columns: anchor and positive
• Approximate statistics based on the first 100 samples:

  	anchor	positive
  type	string	string
  modality	text	text
  details	min: 4 tokens mean: 6.93 tokens max: 14 tokens	min: 5 tokens mean: 8.36 tokens max: 17 tokens

• Samples:

  anchor	positive
  Ultifix LB 200 mg	Cefixime + Lactobacillus
  Rini	Tremoxyl
  Uripro	Prostado 0.4mg

• Loss: CachedGISTEmbedLoss with these parameters:

  {
      "guide": "SentenceTransformer(None)",
      "temperature": 0.01,
      "mini_batch_size": 64,
      "mini_batch_num_tokens": null,
      "margin_strategy": "absolute",
      "margin": 0.0,
      "contrast_anchors": true,
      "contrast_positives": true,
      "gather_across_devices": false
  }
  

triplets

• Dataset: triplets
• Size: 218,643 training samples
• Columns: anchor, positive, and negative
• Approximate statistics based on the first 100 samples:

  	anchor	positive	negative
  type	string	string	string
  modality	text	text	text
  details	min: 4 tokens mean: 7.38 tokens max: 21 tokens	min: 4 tokens mean: 7.69 tokens max: 14 tokens	min: 4 tokens mean: 8.82 tokens max: 14 tokens

• Samples:

  anchor	positive	negative
  Dutasteride + Silodosin	Hakosil-D 8	Dutacet 0.5mg
  Domezole 10mg/20mg	Oara D 10mg/20mg	Ranidom O NF 30 mg/40 mg
  Telmiwell	Teltab 80mg	Inzit TL 80

• Loss: CachedGISTEmbedLoss with these parameters:

  {
      "guide": "SentenceTransformer(None)",
      "temperature": 0.01,
      "mini_batch_size": 64,
      "mini_batch_num_tokens": null,
      "margin_strategy": "absolute",
      "margin": 0.0,
      "contrast_anchors": true,
      "contrast_positives": true,
      "gather_across_devices": false
  }
  

Training Hyperparameters

Non-Default Hyperparameters

• per_device_train_batch_size: 512
• num_train_epochs: 2.0
• learning_rate: 1e-05
• warmup_steps: 0.1
• bf16: True
• load_best_model_at_end: True
• batch_sampler: no_duplicates

All Hyperparameters

Click to expand

• per_device_train_batch_size: 512
• num_train_epochs: 2.0
• max_steps: -1
• learning_rate: 1e-05
• lr_scheduler_type: linear
• 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: 8
• 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: 42
• 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
• 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
• warmup_ratio: None
• local_rank: -1
• prompts: None
• batch_sampler: no_duplicates
• multi_dataset_batch_sampler: proportional
• router_mapping: {}
• learning_rate_mapping: {}

Training Logs

Epoch	Step	Training Loss	comp_nn_accuracy	val_cosine_accuracy	lookalike_cosine_accuracy	unseen_comp_cosine_accuracy
-1	-1	-	0.2394	0.3322	0.1685	0.2690
0.0433	50	8.7900	-	-	-	-
0.0866	100	6.7092	-	-	-	-
0.1299	150	5.1528	0.3697	0.3940	0.2300	0.3223
0.1732	200	4.1212	-	-	-	-
0.2165	250	3.7974	-	-	-	-
0.2597	300	3.5863	0.4296	0.4680	0.3235	0.3895
0.3030	350	3.3763	-	-	-	-
0.3463	400	3.4056	-	-	-	-
0.3896	450	3.2757	0.4577	0.4830	0.3445	0.4072
0.4329	500	3.2090	-	-	-	-
0.4762	550	3.1691	-	-	-	-
0.5195	600	3.1815	0.4542	0.4997	0.3605	0.4125
0.5628	650	3.0593	-	-	-	-
0.6061	700	3.0071	-	-	-	-
0.6494	750	3.0571	0.4683	0.5173	0.3805	0.4272
0.6926	800	2.9532	-	-	-	-
0.7359	850	2.9679	-	-	-	-
0.7792	900	2.9159	0.4718	0.5265	0.3927	0.4442
0.8225	950	3.0154	-	-	-	-
0.8658	1000	2.9073	-	-	-	-
0.9091	1050	1.7601	0.4789	0.5272	0.3940	0.4390
0.9524	1100	1.1786	-	-	-	-
0.9957	1150	1.2053	-	-	-	-
1.0390	1200	2.7288	0.4683	0.5240	0.3842	0.4335
1.0823	1250	2.7543	-	-	-	-
1.1255	1300	2.7919	-	-	-	-
1.1688	1350	2.7564	0.4789	0.5305	0.3965	0.4415
1.2121	1400	2.8945	-	-	-	-
1.2554	1450	2.8302	-	-	-	-
1.2987	1500	2.7203	0.4683	0.5297	0.3963	0.4420
1.3420	1550	2.7254	-	-	-	-
1.3853	1600	2.6467	-	-	-	-
1.4286	1650	2.6925	0.4789	0.5297	0.3952	0.4420
1.4719	1700	2.7150	-	-	-	-
1.5152	1750	2.7332	-	-	-	-
1.5584	1800	2.6659	0.4824	0.5310	0.3947	0.4397
1.6017	1850	2.7103	-	-	-	-
1.6450	1900	2.5801	-	-	-	-
1.6883	1950	2.6788	0.4894	0.5235	0.387	0.4333
1.7316	2000	2.6511	-	-	-	-
1.7749	2050	2.6337	-	-	-	-
1.8182	2100	2.6369	0.4754	0.5265	0.3910	0.4375
1.8615	2150	2.7015	-	-	-	-
1.9048	2200	1.7819	-	-	-	-
1.9481	2250	1.3121	0.4859	0.5320	0.3963	0.4420
1.9913	2300	1.0570	-	-	-	-
2.0	2310	-	0.4789	0.5330	0.3970	0.4440

• The bold row denotes the saved checkpoint.

Training Time

• Training: 43.2 minutes
• Evaluation: 35.9 seconds
• Total: 43.8 minutes

Framework Versions

• Python: 3.12.10
• Sentence Transformers: 5.7.0
• Transformers: 5.14.1
• PyTorch: 2.11.0+cu128
• Accelerate: 1.14.0
• Datasets: 5.0.1
• Tokenizers: 0.22.2

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",
}