MendelAI/nv-embed-v2-ontada-twab-peft

SentenceTransformer based on nvidia/NV-Embed-v2

This is a sentence-transformers model finetuned from nvidia/NV-Embed-v2. It maps sentences & paragraphs to a 4096-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: nvidia/NV-Embed-v2
• Maximum Sequence Length: 1024 tokens
• Output Dimensionality: 4096 dimensions
• Similarity Function: Cosine Similarity

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': 1024, 'do_lower_case': False}) with Transformer model: NVEmbedModel 
  (1): Pooling({'word_embedding_dimension': 4096, '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': False})
  (2): Normalize()
)

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("MendelAI/nv-embed-v2-ontada-twab-peft")
# Run inference
sentences = [
    'Instruct: Given a question, retrieve passages that answer the question. Query: what is the total dose administered in the EBRT Intensity Modulated Radiation Therapy?',
    'Source: SOAP_Note. Date: 2020-03-13. Context: MV electrons.\n \n FIELDS:\n The right orbital mass and right cervical lymph nodes were initially treated with a two arc IMRT plan. Arc 1: 11.4 x 21 cm. Gantry start and stop angles 178 degrees / 182 degrees. Arc 2: 16.4 x 13.0 cm. Gantry start ',
    'Source: Radiology. Date: 2023-09-18. Context: : >60\n \n Contrast Type: OMNI 350\n   Volume: 80ML\n \n Lot_: ________\n \n Exp. date: 05/26 \n Study Completed: CT CHEST W\n \n Reading Group:BCH \n  \n   Prior Studies for Comparison: 06/14/23 CT CHEST W RMCC  \n \n ________ ______\n  ',
]
embeddings = model.encode(sentences)
print(embeddings.shape)
# [3, 4096]

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

Evaluation

Metrics

Patient QA

• Dataset: ontada-test
• Evaluated with PatientQAEvaluator

Metric	Value
cosine_accuracy@1	0.6856
cosine_accuracy@3	0.9531
cosine_accuracy@5	0.9909
cosine_accuracy@10	1.0
cosine_precision@1	0.6856
cosine_precision@3	0.5209
cosine_precision@5	0.3969
cosine_precision@10	0.2251
cosine_recall@1	0.4203
cosine_recall@3	0.8154
cosine_recall@5	0.9454
cosine_recall@10	1.0046
cosine_ndcg@10	0.8649
cosine_mrr@10	0.8191
cosine_map@100	0.805

Training Details

Training Dataset

Unnamed Dataset

• Size: 16,186 training samples
• Columns: question and context
• Approximate statistics based on the first 1000 samples:

  	question	context
  type	string	string
  details	min: 25 tokens mean: 30.78 tokens max: 39 tokens	min: 74 tokens mean: 177.84 tokens max: 398 tokens

• Samples:

  question	context
  Instruct: Given a question, retrieve passages that answer the question. Query: what was the abnormality identified for BRAF?	Source: Genetic_Testing. Date: 2022-10-07. Context: Mutational Seq DNA-Tumor Low, 6 mt/Mb NF1 Seq DNA-Tumor Mutation Not Detected T In Not D ARID2 Seq DNA-Tumor Mutation Not Detected CNA-Seq DNA-Tumor Deletion Not Detected PTEN Seq RNA-Tumor Fusion Not Detected Seq DNA-Tumor Mutation Not Detected BRAF Amplification Not _ CNA-Seq DNA-Tumor Detected RAC1 Seq DNA-Tumor Mutation Not Detected The selection of any, all, or none of the matched therapies
  Instruct: Given a question, retrieve passages that answer the question. Query: what was the abnormality identified for BRAF?	Source: Genetic_Testing. Date: 2021-06-04. Context: characteristics have been determined by _____ ____ _________ ___ ____ _______. It has not been cleared or approved by FDA. This assay has been validated pursuant to the CLIA regulations and is used for clinical purposes. BRAF MUTATION ANALYSIS E SOURCE: LYMPH NODE PARAFFIN BLOCK NUMBER: - A4 BRAF MUTATION ANALYSIS NOT DETECTED NOT DETECTED This result was reviewed and interpreted by _. ____, M.D. Based on Sanger sequencing analysis, no mutations
  Instruct: Given a question, retrieve passages that answer the question. Query: what was the abnormality identified for BRAF?	Source: Pathology. Date: 2019-12-12. Context: Receive Date: 12/12/2019 ___ _: ________________ Accession Date: 12/12/2019 Copy To: Report Date: 12/19/2019 18:16 SUPPLEMENTAL REPORT (previous report date: 12/19/2019) BRAF SNAPSHOT Results: POSITIVE Interpretation: A BRAF mutation was detected in the provided specimen. FDA has approved TKI inhibitor vemurafenib and dabrafenib for the first-line treatment of patients with unresectable or metastatic melanoma whose tumors have a BRAF V600E mutation, and trametinib for tumors

• Loss: MultipleNegativesRankingLoss with these parameters:

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

Training Hyperparameters

Non-Default Hyperparameters

• eval_strategy: steps
• per_device_train_batch_size: 4
• per_device_eval_batch_size: 64
• learning_rate: 2e-05
• num_train_epochs: 1
• warmup_ratio: 0.1
• seed: 6789
• bf16: True
• prompts: {'question': 'Instruct: Given a question, retrieve passages that answer the question. Query: '}
• batch_sampler: no_duplicates

All Hyperparameters

Click to expand

• overwrite_output_dir: False
• do_predict: False
• eval_strategy: steps
• prediction_loss_only: True
• per_device_train_batch_size: 4
• per_device_eval_batch_size: 64
• 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: 6789
• data_seed: None
• jit_mode_eval: False
• use_ipex: False
• bf16: True
• fp16: False
• 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
• 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
• 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
• use_liger_kernel: False
• eval_use_gather_object: False
• prompts: {'question': 'Instruct: Given a question, retrieve passages that answer the question. Query: '}
• batch_sampler: no_duplicates
• multi_dataset_batch_sampler: proportional

Training Logs

Epoch	Step	Training Loss	ontada-test_cosine_ndcg@10
0	0	-	0.8431
0.0002	1	1.5826	-
0.0371	150	0.4123	-
0.0741	300	0.3077	-
0.1112	450	0.2184	-
0.1483	600	0.3291	-
0.1853	750	0.2343	-
0.2224	900	0.2506	-
0.2471	1000	-	0.8077
0.2595	1050	0.1294	-
0.2965	1200	0.0158	-
0.3336	1350	0.0189	-
0.3706	1500	0.0363	-
0.4077	1650	0.0208	-
0.4448	1800	0.475	-
0.4818	1950	0.6183	-
0.4942	2000	-	0.8482
0.5189	2100	0.4779	-
0.5560	2250	0.4194	-
0.5930	2400	0.8376	-
0.6301	2550	0.4249	-
0.6672	2700	0.9336	-
0.7042	2850	0.5351	-
0.7413	3000	1.0253	0.8551
0.7784	3150	0.3961	-
0.8154	3300	0.3881	-
0.8525	3450	0.5573	-
0.8895	3600	1.222	-
0.9266	3750	0.3032	-
0.9637	3900	0.3142	-
0.9884	4000	-	0.8645
1.0	4047	-	0.8649

Framework Versions

• Python: 3.11.10
• Sentence Transformers: 3.4.0.dev0
• Transformers: 4.46.0
• PyTorch: 2.3.1+cu121
• Accelerate: 1.0.1
• Datasets: 3.0.1
• Tokenizers: 0.20.1

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}
}