wwydmanski/specter2_pubmed-v0.5

SentenceTransformer based on allenai/specter2_base

This model is an initial proof of concept for (yet unpublished) article on ultra-hard negative triplet generation. While the original Specter2 adapters were trained on 600k triplets, only 10k ultra-hard, self-supervised negatives were enough to outperform the Proximity adapter (85 vs 84.1 avg NDCG over Relish, NFCorpus, TREC CoVID).

Model Details

This is a sentence-transformers model finetuned from allenai/specter2_base on the json 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 Description

• Model Type: Sentence Transformer
• Base model: allenai/specter2_base
• Maximum Sequence Length: 512 tokens
• Output Dimensionality: 768 tokens
• Similarity Function: Cosine Similarity
• Training Dataset:
  • json

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: BertModel 
  (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("sentence_transformers_model_id")
# Run inference
sentences = [
    'Six-coordinate low-spin iron(III) porphyrinate complexes',
    'Molecular structures and magnetic resonance spectroscopic investigations of highly distorted six-coordinate low-spin iron(III) porphyrinate complexes. ',
    'Saddle-shaped six-coordinate iron(iii) porphyrin complex with unusual intermediate-spin electronic structure. ',
]
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: triplet-dev
• Evaluated with TripletEvaluator

Metric	Value
cosine_accuracy	0.606
dot_accuracy	0.395
manhattan_accuracy	0.603
euclidean_accuracy	0.615
max_accuracy	0.615

Training Details

Training Dataset

json

• Dataset: json
• Size: 10,053 training samples
• Columns: anchor, positive, and negative
• Approximate statistics based on the first 1000 samples:

  	anchor	positive	negative
  type	string	string	string
  details	min: 4 tokens mean: 7.49 tokens max: 18 tokens	min: 4 tokens mean: 20.08 tokens max: 48 tokens	min: 4 tokens mean: 12.46 tokens max: 48 tokens

• Samples:

  anchor	positive	negative
  COM-induced secretome changes in U937 monocytes	Characterization of calcium oxalate crystal-induced changes in the secretome of U937 human monocytes.	Monocytes.
  Metamaterials	Sound attenuation optimization using metaporous materials tuned on exceptional points.	Metamaterials: A cat's eye for all directions.
  Pediatric Parasitology	Parasitic infections among school age children 6 to 11-years-of-age in the Eastern province.	[DIALOGUE ON PEDIATRIC PARASITOLOGY].

• 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: 32
• per_device_eval_batch_size: 32
• learning_rate: 2e-05
• num_train_epochs: 6
• lr_scheduler_type: cosine_with_restarts
• warmup_ratio: 0.1
• bf16: True
• 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: 32
• per_device_eval_batch_size: 32
• 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: 6
• max_steps: -1
• lr_scheduler_type: cosine_with_restarts
• 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: 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
• 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
• batch_sampler: no_duplicates
• multi_dataset_batch_sampler: proportional

Training Logs

Epoch	Step	Training Loss	triplet-dev_cosine_accuracy
0	0	-	0.373
0.1667	1	3.138	-
0.3333	2	2.9761	-
0.5	3	2.7135	-
0.6667	4	2.5144	-
0.8333	5	1.9797	-
1.0	6	1.2683	-
1.1667	7	1.6058	-
1.3333	8	1.3236	-
1.5	9	1.1134	-
1.6667	10	1.1205	-
1.8333	11	0.9369	-
2.0	12	0.6215	-
2.1667	13	1.0374	-
2.3333	14	0.9355	-
2.5	15	0.7118	-
2.6667	16	0.7967	-
2.8333	17	0.5739	-
3.0	18	0.4515	-
3.1667	19	0.8018	-
3.3333	20	0.6557	-
3.5	21	0.6027	-
3.6667	22	0.6747	-
3.8333	23	0.5013	-
4.0	24	0.1428	-
4.1667	25	0.5889	0.596
4.3333	26	0.5439	-
4.5	27	0.4742	-
4.6667	28	0.5734	-
4.8333	29	0.3966	-
5.0	30	0.1793	-
5.1667	31	0.5408	-
5.3333	32	0.5174	-
5.5	33	0.4179	-
5.6667	34	0.4589	-
5.8333	35	0.3683	-
6.0	36	0.1442	0.606

Framework Versions

• Python: 3.9.19
• Sentence Transformers: 3.1.1
• Transformers: 4.45.2
• PyTorch: 2.5.0
• Accelerate: 1.0.1
• Datasets: 2.19.0
• Tokenizers: 0.20.3

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}
}