lsy9874205/heal-protocol-embeddings

SentenceTransformer based on sentence-transformers/all-MiniLM-L6-v2

This is a sentence-transformers model finetuned from sentence-transformers/all-MiniLM-L6-v2. It maps sentences & paragraphs to a 384-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: sentence-transformers/all-MiniLM-L6-v2
• Maximum Sequence Length: 256 tokens
• Output Dimensionality: 384 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': 256, 'do_lower_case': False}) with Transformer model: BertModel 
  (1): Pooling({'word_embedding_dimension': 384, '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 (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 = [
    'To evaluate the cost-effectiveness of ED-initiated buprenorphine with peer navigator support compared to enhanced referral to treatment.',
    'Concerns about withdrawal precipitation',
    '11.1.2 Steering Committee\n\nComposition:\n- Executive Committee members\n- Site investigators\n- Patient/community representatives\n- Key co-investigators\n\nResponsibilities:\n- Protocol revisions\n- Implementation monitoring\n- Recruitment oversight\n- Review of study progress\n- Addressing operational challenges',
]
embeddings = model.encode(sentences)
print(embeddings.shape)
# [3, 384]

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

Training Details

Training Dataset

Unnamed Dataset

• Size: 247,936 training samples
• Columns: sentence_0, sentence_1, and label
• Approximate statistics based on the first 1000 samples:

  	sentence_0	sentence_1	label
  type	string	string	float
  details	min: 3 tokens mean: 48.16 tokens max: 256 tokens	min: 3 tokens mean: 44.76 tokens max: 256 tokens	min: 0.5 mean: 0.5 max: 1.0

• Samples:

  sentence_0	sentence_1	label
  10.4 Participant Confidentiality	9.1.1 Data and Safety Monitoring Board (DSMB) An independent DSMB will be established, consisting of experts in emergency medicine, addiction medicine, biostatistics, and ethics. The DSMB will: - Review and approve the monitoring plan - Meet at least annually to review study progress and safety - Review any serious adverse events - Make recommendations regarding study continuation or modification	0.5
  7.1 Randomization Participants will be randomly assigned in a 1:1 ratio to receive either BUP-NX or XR-NTX using a computer-generated randomization sequence with permuted blocks of varying sizes. Randomization will be stratified by site and by opioid type (short-acting prescription opioids, heroin, or fentanyl as primary opioid of use).	10.3 Risk Mitigation	0.5
  11.1 Study Leadership and Governance	To examine patient perspectives on intervention acceptability and barriers/facilitators to engagement through qualitative interviews with a subset of participants.	0.5

• Loss: CosineSimilarityLoss with these parameters:

  {
      "loss_fct": "torch.nn.modules.loss.MSELoss"
  }
  

Training Hyperparameters

Non-Default Hyperparameters

• per_device_train_batch_size: 16
• per_device_eval_batch_size: 16
• multi_dataset_batch_sampler: round_robin

All Hyperparameters

Click to expand

• overwrite_output_dir: False
• do_predict: False
• eval_strategy: no
• prediction_loss_only: True
• per_device_train_batch_size: 16
• per_device_eval_batch_size: 16
• 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: 5e-05
• weight_decay: 0.0
• adam_beta1: 0.9
• adam_beta2: 0.999
• adam_epsilon: 1e-08
• max_grad_norm: 1
• num_train_epochs: 3
• max_steps: -1
• lr_scheduler_type: linear
• lr_scheduler_kwargs: {}
• warmup_ratio: 0.0
• 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: 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: None
• 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
• average_tokens_across_devices: False
• prompts: None
• batch_sampler: batch_sampler
• multi_dataset_batch_sampler: round_robin

Training Logs

Epoch	Step	Training Loss
0.0323	500	0.0107
0.0645	1000	0.0025
0.0968	1500	0.0023
0.1291	2000	0.0023
0.1613	2500	0.0021
0.1936	3000	0.002
0.2259	3500	0.0018
0.2581	4000	0.0018
0.2904	4500	0.0017
0.3227	5000	0.0017
0.3549	5500	0.0017
0.3872	6000	0.0016
0.4195	6500	0.0015
0.4517	7000	0.0016
0.4840	7500	0.0016
0.5163	8000	0.0015
0.5485	8500	0.0015
0.5808	9000	0.0014
0.6131	9500	0.0015
0.6453	10000	0.0015
0.6776	10500	0.0014
0.7099	11000	0.0015
0.7421	11500	0.0013
0.7744	12000	0.0013
0.8067	12500	0.0013
0.8389	13000	0.0013
0.8712	13500	0.0013
0.9035	14000	0.0013
0.9357	14500	0.0013
0.9680	15000	0.0012
1.0003	15500	0.0012
1.0325	16000	0.0011
1.0648	16500	0.0011
1.0971	17000	0.0011
1.1293	17500	0.0011
1.1616	18000	0.0011
1.1939	18500	0.001
1.2261	19000	0.001
1.2584	19500	0.0011
1.2907	20000	0.001
1.3229	20500	0.0011
1.3552	21000	0.001
1.3875	21500	0.001
1.4197	22000	0.001
1.4520	22500	0.001
1.4843	23000	0.001
1.5165	23500	0.0009
1.5488	24000	0.001
1.5811	24500	0.001
1.6133	25000	0.0009
1.6456	25500	0.001
1.6779	26000	0.001
1.7101	26500	0.001
1.7424	27000	0.001
1.7747	27500	0.001
1.8069	28000	0.001
1.8392	28500	0.001
1.8715	29000	0.001
1.9037	29500	0.0009
1.9360	30000	0.0009
1.9682	30500	0.0009
2.0005	31000	0.0009
2.0328	31500	0.0008
2.0650	32000	0.0008
2.0973	32500	0.0007
2.1296	33000	0.0008
2.1618	33500	0.0008
2.1941	34000	0.0008
2.2264	34500	0.0008
2.2586	35000	0.0008
2.2909	35500	0.0008
2.3232	36000	0.0008
2.3554	36500	0.0008
2.3877	37000	0.0008
2.4200	37500	0.0008
2.4522	38000	0.0008
2.4845	38500	0.0008
2.5168	39000	0.0008
2.5490	39500	0.0008
2.5813	40000	0.0007
2.6136	40500	0.0008
2.6458	41000	0.0008
2.6781	41500	0.0007
2.7104	42000	0.0007
2.7426	42500	0.0007
2.7749	43000	0.0008
2.8072	43500	0.0008
2.8394	44000	0.0007
2.8717	44500	0.0008
2.9040	45000	0.0008
2.9362	45500	0.0007
2.9685	46000	0.0007

Framework Versions

• Python: 3.13.2
• Sentence Transformers: 3.4.1
• Transformers: 4.49.0
• PyTorch: 2.6.0
• Accelerate: 1.4.0
• Datasets: 3.3.2
• Tokenizers: 0.21.0

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",
}

HEAL Protocol Embeddings

This model is fine-tuned from all-MiniLM-L6-v2 on HEAL Initiative clinical protocols.

Performance Evaluation

Comparison with OpenAI embeddings:

Metric	OpenAI	Fine-tuned	Change
Faithfulness	0.667	0.833	⬆️ +0.166
Answer Relevancy	0.986	0.986	=
Context Precision	1.000	1.000	=
Context Recall	1.000	0.000	⬇️ -1.000

Key Findings

• Improved faithfulness to source material
• Maintained high answer relevancy
• Trade-off in context recall

Future Improvements

1. Retrieval Strategy

   • Implement hybrid search combining semantic and keyword matching
   • Add re-ranking for better result ordering

2. Model Architecture

   • Experiment with larger base models
   • Fine-tune with domain-specific loss functions

3. Data Processing

   • Optimize chunking strategy
   • Increase training data diversity