SentenceTransformer based on Alibaba-NLP/gte-modernbert-base

This is a sentence-transformers model finetuned from Alibaba-NLP/gte-modernbert-base. 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: Alibaba-NLP/gte-modernbert-base
Maximum Sequence Length: 8192 tokens
Output Dimensionality: 768 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': 8192, 'do_lower_case': False}) with Transformer model: ModernBertModel 
  (1): Pooling({'word_embedding_dimension': 768, 'pooling_mode_cls_token': True, 'pooling_mode_mean_tokens': False, '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("anasse15/MNLP_M3_document_encoder")
# Run inference
sentences = [
    "What is the primary role of davemaoite in Earth's lower mantle?\nA. It is the most abundant mineral in the crust.\nB. It acts as a catalyst for mineral formation.\nC. It serves as a primary source of diamonds.\nD. It contributes to heat flow through radioactive decay.",
    "Davemaoite  is  a high-pressure calcium silicate perovskite (CaSiO3) mineral with a distinctive cubic crystal structure. It is named after geophysicist Ho-kwang (Dave) Mao, who pioneered in many discoveries in high-pressure geochemistry and geophysics.  \n\nIt is one of three main minerals in Earth’s lower mantle, making up around 5–7% of the material there. Significantly, davemaoite can host uranium and thorium, radioactive isotopes which produce heat through radioactive decay and contribute greatly to heating within this region giving the material a major role in how heat flows deep below the earth's surface.\n\nDavemaoite has been artificially synthesized in the laboratory, but was thought to be too extreme to exist in the Earth's crust. Then in 2021, the mineral was discovered as specks within a diamond that formed between 660 and 900 km beneath the Earth's surface, within the mantle. The diamond had been extracted from the Orapa diamond mine in Botswana. The discovery was made  by focusing a high-energy beam of X-rays on precise spots within the diamond  using a technique known as synchrotron X-ray diffraction. \n\nCalcium silicate is found in other forms, such as wollastonite in the crust and breyite in the middle and lower regions of the mantle. However, this version can exist only at very high pressure of around 200,000 times that found at Earth’s surface.\n\nSee also\n\n Perovskite (structure)\nList of minerals\n\nReferences \n\nPerovskites\nCalcium minerals",
    'In molecular biology, the calcipressin family of proteins negatively regulate calcineurin by direct binding. They are essential for the survival of T helper type 1 cells. Calcipressin 1 is a phosphoprotein that increases its capacity to inhibit calcineurin when phosphorylated at the conserved FLISPP motif; this phosphorylation also controls the half-life of calcipressin 1 by accelerating its degradation.\n\nIn humans, the Calcipressins family of proteins is derived from three genes. Calcipressin 1 is also known as modulatory calcineurin-interacting protein 1 (MCIP1), Adapt78 and Down syndrome critical region 1 (DSCR1). Calcipressin 2 is variously known as MCIP2, ZAKI-4 and DSCR1-like 1. Calcipressin 3 is also called MCIP3 and DSCR1-like 2.\n\nReferences\n\nProtein families',
]
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: validation
Evaluated with TripletEvaluator

Metric	Value
cosine_accuracy	1.0

Training Details

Training Dataset

Unnamed Dataset

Size: 12,689 training samples
Columns: sentence_0, sentence_1, and sentence_2

Approximate statistics based on the first 1000 samples:

	sentence_0	sentence_1	sentence_2
type	string	string	string
details	min: 30 tokens mean: 84.52 tokens max: 198 tokens	min: 94 tokens mean: 261.34 tokens max: 818 tokens	min: 101 tokens mean: 257.86 tokens max: 752 tokens

Samples:

sentence_0	sentence_1	sentence_2
`What type of model is the TaiWan Ionospheric Model (TWIM)? A. A one-dimensional thermal model of the Earth's crust B. A two-dimensional statistical model of atmospheric pressure C. A four-dimensional quantum model of particle interactions D. A three-dimensional numerical and phenomenological model of ionospheric electron density`	The TaiWan Ionospheric Model (TWIM) developed in 2008 is a three-dimensional numerical and phenomenological model of ionospheric electron density (Ne). The TWIM has been constructed from global distributed ionosonde foF2 and foE data and vertical Ne profiles retrieved from FormoSat3/COSMIC GPS radio occultation measurements. The TWIM consists of vertically fitted α-Chapman-type layers, with distinct F2, F1, E, and D layers, for which the layer parameters such as peak density, peak density height, and scale height are represented by surface spherical harmonics. These results are useful for providing reliable radio propagation predictions and in investigation of near-Earth space and large-scale Ne distribution with diurnal and seasonal variations, along with geographic features such as the equatorial anomaly. This way the continuity of Ne and its derivatives is also maintained for practical schemes for providing reliable radio propagation predictions. References The information in thi...	Chandrasekhar–Kendall functions are the axisymmetric eigenfunctions of the curl operator, derived by Subrahmanyan Chandrasekhar and P.C. Kendall in 1957, in attempting to solve the force-free magnetic fields. The results were independently derived by both, but were agreed to publish the paper together. If the force-free magnetic field equation is written as with the assumption of divergence free field (), then the most general solution for axisymmetric case is where is a unit vector and the scalar function satisfies the Helmholtz equation, i.e., The same equation also appears in fluid dynamics in Beltrami flows where, vorticity vector is parallel to the velocity vector, i.e., . Derivation Taking curl of the equation and using this same equation, we get . In the vector identity , we can set since it is solenoidal, which leads to a vector Helmholtz equation, . Every solution of above equation is not the solution of original equation, but the converse is true. If is a scal...
`What is the primary function of the protein encoded by the PFN2 gene? A. Facilitating lipid metabolism B. Regulating actin polymerization C. Encoding DNA repair enzymes D. Transporting oxygen in blood`	Profilin-2 is a protein that in humans is encoded by the PFN2 gene. The protein encoded by this gene is a ubiquitous actin monomer-binding protein belonging to the profilin family. It is thought to regulate actin polymerization in response to extracellular signals. There are two alternatively spliced transcript variants encoding different isoforms described for this gene. Interactions PFN2 has been shown to interact with ROCK1, Vasodilator-stimulated phosphoprotein, CCDC113 and FMNL1. References Further reading External links	`Stearoyl-CoA is a coenzyme involved in the metabolism of fatty acids. Stearoyl-CoA is an 18-carbon long fatty acyl-CoA chain that participates in an unsaturation reaction. The reaction is catalyzed by the enzyme stearoyl-CoA desaturase, which is located in the endoplasmic reticulum. It forms a cis-double bond between the ninth and tenth carbons within the chain to form the product oleoyl-CoA. References Bibliography Metabolism Thioesters of coenzyme A`
`Which of the following statements is true regarding the properties of certain mathematical spaces and their relevance in functional analysis? A. Souslin spaces are always separable and complete metrizable. B. All Polish spaces are K-analytic but not all K-analytic spaces are Polish. C. The Borel graph theorem applies only to finite-dimensional spaces. D. The VEZF1 gene is involved in the continuity of linear maps in functional analysis.`	`Vascular endothelial zinc finger 1 is a protein that in humans is encoded by the VEZF1 gene. Function Transcriptional regulatory proteins containing tandemly repeated zinc finger domains are thought to be involved in both normal and abnormal cellular proliferation and differentiation. ZNF161 is a C2H2-type zinc finger protein (Koyano-Nakagawa et al., 1994 [PubMed 8035792]). See MIM 603971 for general information on zinc finger proteins. References Further reading`	`In mathematics, a trivial semigroup (a semigroup with one element) is a semigroup for which the cardinality of the underlying set is one. The number of distinct nonisomorphic semigroups with one element is one. If S = { a } is a semigroup with one element, then the Cayley table of S is {`

Loss: main.TripletLossWithLogging with these parameters:

{
    "distance_metric": "TripletDistanceMetric.EUCLIDEAN",
    "triplet_margin": 5
}

Training Hyperparameters

Non-Default Hyperparameters

eval_strategy: steps
per_device_train_batch_size: 16
per_device_eval_batch_size: 16
num_train_epochs: 1
fp16: True
multi_dataset_batch_sampler: round_robin

All Hyperparameters

Click to expand

overwrite_output_dir: False
do_predict: False
eval_strategy: steps
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: 1
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: 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: 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
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	validation_cosine_accuracy
0.1259	100	-	1.0
0.2519	200	-	1.0
0.3778	300	-	1.0
0.5038	400	-	1.0
0.6297	500	0.1864	1.0
0.7557	600	-	1.0
0.8816	700	-	1.0
1.0	794	-	1.0

Framework Versions

Python: 3.12.8
Sentence Transformers: 4.1.0
Transformers: 4.52.3
PyTorch: 2.7.0+cu126
Accelerate: 1.3.0
Datasets: 3.6.0
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",
}

TripletLossWithLogging

@misc{hermans2017defense,
    title={In Defense of the Triplet Loss for Person Re-Identification},
    author={Alexander Hermans and Lucas Beyer and Bastian Leibe},
    year={2017},
    eprint={1703.07737},
    archivePrefix={arXiv},
    primaryClass={cs.CV}
}

anasse15
/

MNLP_M3_document_encoder