This is a sentence-transformers model finetuned from BAAI/bge-m3 on the tqa-bge-m3-finetune-triplets dataset. It maps sentences & paragraphs to a 1024-dimensional dense vector space and can be used for retrieval.

Model Details

Model Description

Model Type: Sentence Transformer
Base model: BAAI/bge-m3
Maximum Sequence Length: 512 tokens
Output Dimensionality: 1024 dimensions
Similarity Function: Cosine Similarity
Supported Modality: Text
Training Dataset:
- tqa-bge-m3-finetune-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': 'XLMRobertaModel'})
  (1): Pooling({'embedding_dimension': 1024, 'pooling_mode': 'cls', '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.

python

from sentence_transformers import SentenceTransformer

# Download from the 🤗 Hub
model = SentenceTransformer("TesNik369/bge-m3-tqa-merged-epoch-3")
# Run inference
queries = [
    'A map in which all the lines of latitude and longitude are straight lines is a',
]
documents = [
    'Most maps use a grid of lines to help you to find your location. This grid system is called a geographic coordinate system. Using this system you can define your location by two numbers, latitude and longitude. Both numbers are angles between your location, the center of Earth, and a reference line (Figure 2.20).',
    'Topographic maps represent the locations of geographical features, such as hills and valleys. Topographic maps use contour lines to show different elevations. A contour line is a line of equal elevation. If you walk along a contour line you will not go uphill or downhill. Topographic maps are also called contour maps. The rules of topographic maps are: Each line connects all points of a specific elevation. Contour lines never cross since a single point can only have one elevation. Every fifth contour line is bolded and labeled. Adjacent contour lines are separated by a constant difference in elevation (such as 20 ft or 100 ft). The difference in elevation is the contour interval, which is indicated in the map legend. Scales indicate horizontal distance and are also found on the map legend. Old Faithful erupting, Yellowstone Na- tional Park. While the Figure 1.1 isnt exactly the same view as the map at the top of this concept, it is easy to see the main features. Hills, forests, development, and trees are all seen around Old Faithful.',
    'In addition to sunlight, aquatic producers also need dissolved oxygen and nutrients. Water near the surface generally contains more dissolved oxygen than deeper water. Many nutrients enter the water from the land. Therefore, water nearer shore usually contains more dissolved nutrients than water farther from shore.',
]
query_embeddings = model.encode_query(queries)
document_embeddings = model.encode_document(documents)
print(query_embeddings.shape, document_embeddings.shape)
# [1, 1024] [3, 1024]

# Get the similarity scores for the embeddings
similarities = model.similarity(query_embeddings, document_embeddings)
print(similarities)
# tensor([[0.5938, 0.6104, 0.2369]])

Evaluation

Metrics

Information Retrieval

Dataset: tqa-dev
Evaluated with InformationRetrievalEvaluator

Metric	Value
cosine_accuracy@1	0.4319
cosine_accuracy@3	0.7621
cosine_accuracy@5	0.8776
cosine_accuracy@10	0.9515
cosine_precision@1	0.4319
cosine_precision@3	0.254
cosine_precision@5	0.1755
cosine_precision@10	0.0952
cosine_recall@1	0.4319
cosine_recall@3	0.7621
cosine_recall@5	0.8776
cosine_recall@10	0.9515
cosine_ndcg@10	0.6987
cosine_mrr@10	0.6165
cosine_map@100	0.6189

Training Details

Training Dataset

tqa-bge-m3-finetune-triplets

Dataset: tqa-bge-m3-finetune-triplets
Size: 8,220 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: 5 tokens mean: 15.4 tokens max: 28 tokens	min: 15 tokens mean: 197.27 tokens max: 512 tokens	min: 9 tokens mean: 189.78 tokens max: 512 tokens

Samples:

anchor	positive	negative
`maximum distance the particles of a medium move from their resting position`	`The speed of most waves depends on the medium through which they are traveling. Generally, waves travel fastest through solids and slowest through gases. Thats because particles are closest together in solids and farthest apart in gases. When particles are farther apart, it takes longer for the energy of the disturbance to pass from particle to particle.`	`The speed of most waves depends on the medium, or the matter through which the waves are traveling. Generally, waves travel fastest through solids and slowest through gases. Thats because particles are closest together in solids and farthest apart in gases. When particles are farther apart, it takes longer for the energy of the disturbance to pass from particle to particle through the medium. Click image to the left or use the URL below. URL:`
`One function of the cornea is to protect the eye.`	The structure of the human eye is shown in Figure 22.24. Find each structure in the diagram as you read about it below. The cornea is the transparent outer covering of the eye. It protects the eye and also acts as a convex lens, helping to focus light that enters the eye. The pupil is an opening in the front of the eye. It looks black because it doesnt reflect any light. It allows light to enter the eye. The pupil automatically gets bigger or smaller to let more or less light in as needed. The iris is the colored part of the eye. It controls the size of the pupil. The lens is a convex lens that fine-tunes the focus so an image forms on the back of the eye. Tiny muscles control the shape of the lens to focus images of close or distant objects. The retina is a membrane lining the back of the eye. The retina has nerve cells called rods and cones that change images to electrical signals. Rods are good at sensing dim light but cant distinguish different colors of light. Cones can sense colo...	The function of the eye is to focus light and form images. We see some objects, such as stars and light bulbs, because they give off their own light. However, we see most objects because they reflect light from another source such as the sun. We form images of the objects when some of the reflected light enters our eyes. Look at the parts of the eye in Figure 20.12. Follow the path of light through the eye as you read about it below. 1. Light from an object passes first through the cornea. This is a clear, protective covering on the outside of the eye. 2. Then light passes through the pupil, an opening in the center of the eye. The pupil, which looks black, is surrounded by the colored part of the eye, called the iris. 3. Light entering through the pupil next passes through the lens. The lens is a clear, curved structure, like the lens of a magnifying glass. Along with the cornea, the lens focuses the light on the back of the eye. 4. The back of the eye is covered by a thin layer calle...
`part of the ear that transmits and amplifies vibrations from the eardrum`	The outer ear includes the pinna, ear canal, and eardrum. The pinna is the only part of the ear that extends outward from the head. Its position and shape make it good at catching sound waves and funneling them into the ear canal. The ear canal is a tube that carries sound waves into the ear. The sound waves travel through the air inside the ear canal to the eardrum. The eardrum is like the head of a drum. Its a thin membrane stretched tight across the end of the ear canal. The eardrum vibrates when sound waves strike it, and it sends the vibrations on to the middle ear.	The outer ear includes the pinna, ear canal, and eardrum. The pinna is the only part of the ear that extends outward from the head. Its position and shape make it good at catching sound waves and funneling them into the ear canal. The ear canal is a tube that carries sound waves into the ear. The sound waves travel through the air inside the ear canal to the eardrum. The eardrum is like the head of a drum. It is a thin membrane stretched tight across the end of the ear canal. The eardrum vibrates when sound waves strike it, and it sends the vibrations on to the middle ear. Q: How might cupping his hands behind his ears help the boy pictured in the opening image hear better? A: His hands might help the pinna of his ears catch sound waves and direct them into the ear canal.

Loss: MultipleNegativesRankingLoss with these parameters:
json
1{ 2 "scale": 20.0, 3 "similarity_fct": "cos_sim", 4 "gather_across_devices": false, 5 "directions": [ 6 "query_to_doc" 7 ], 8 "partition_mode": "joint", 9 "hardness_mode": null, 10 "hardness_strength": 0.0 11}

Training Hyperparameters

Non-Default Hyperparameters

per_device_train_batch_size: 16
per_device_eval_batch_size: 32
gradient_accumulation_steps: 8
learning_rate: 2e-05
lr_scheduler_type: cosine
warmup_ratio: 0.05
seed: 3407
fp16: True
gradient_checkpointing: True
batch_sampler: no_duplicates

All Hyperparameters

Click to expand

overwrite_output_dir: False
do_predict: False
prediction_loss_only: True
per_device_train_batch_size: 16
per_device_eval_batch_size: 32
per_gpu_train_batch_size: None
per_gpu_eval_batch_size: None
gradient_accumulation_steps: 8
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: 3
max_steps: -1
lr_scheduler_type: cosine
lr_scheduler_kwargs: {}
warmup_ratio: 0.05
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: 3407
data_seed: None
jit_mode_eval: 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}
parallelism_config: None
deepspeed: None
label_smoothing_factor: 0.0
optim: adamw_torch_fused
optim_args: None
adafactor: False
group_by_length: False
length_column_name: length
project: huggingface
trackio_space_id: trackio
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
hub_revision: None
gradient_checkpointing: True
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: no
neftune_noise_alpha: None
optim_target_modules: None
batch_eval_metrics: False
eval_on_start: False
use_liger_kernel: False
liger_kernel_config: None
eval_use_gather_object: False
average_tokens_across_devices: True
prompts: None
batch_sampler: no_duplicates
multi_dataset_batch_sampler: proportional
router_mapping: {}
learning_rate_mapping: {}

Training Logs

Epoch	Step	Training Loss	tqa-dev_cosine_ndcg@10
0.0156	1	1.129	-
0.3113	20	1.0796	-
0.6226	40	1.047	-
0.9339	60	1.0395	-
1.0	65	-	0.7032
1.2335	80	1.0342	-
1.5447	100	0.9842	-
1.8560	120	0.972	-
2.0	130	-	0.7020
2.1556	140	0.9788	-
2.4669	160	0.9826	-
2.7782	180	0.9577	-
3.0	195	-	0.6987
-1	-1	-	0.6987

Training Time

Training: 1.5 hours

Framework Versions

Python: 3.12.12
Sentence Transformers: 5.6.1
Transformers: 4.57.1
PyTorch: 2.8.0+cu126
Accelerate: 1.11.0
Datasets: 5.0.0
Tokenizers: 0.22.1

Citation

BibTeX

Sentence Transformers

bibtex

@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

bibtex

@misc{oord2019representationlearningcontrastivepredictive,
      title={Representation Learning with Contrastive Predictive Coding},
      author={Aaron van den Oord and Yazhe Li and Oriol Vinyals},
      year={2019},
      eprint={1807.03748},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/1807.03748},
}