Add new SentenceTransformer model

def8c03 verified 3 months ago

21 kB

tags:
  - sentence-transformers
  - sentence-similarity
  - feature-extraction
  - dense
  - generated_from_trainer
  - dataset_size:123502
  - loss:MultipleNegativesRankingLoss
base_model: google/embeddinggemma-300m
widget:
  - source_sentence: стержни клеевые 11 x 200 mm прозрачные
    sentences:
      - шпаклевка danogips superfinish готовая 18 кг
      - стержни для клеевого пистолета biber 60131 прозр. 11 x 200 mm 6 шт.
      - наличник деревянный прямой сращенный сорт экстра 11 x 90 x 2200 mm
  - source_sentence: стамеска 18 mm biber мастер
    sentences:
      - стамеска biber 85057 мастер 18 mm
      - стамеска biber 85053 мастер 10 mm
      - наконечник ншви 2 6 0 14 для провода 6 mm2 100 шт.
  - source_sentence: саморез 3.8 x 41 mm шсгд
    sentences:
      - коробка упаковочная сатурн 550 x 380 x 300 mm картон
      - саморез swfs шсгд 3.8 x 51 200 шт
      - саморез swfs шсгд zy 3.8 x 41 25 шт пакетик
  - source_sentence: гибкое стекло 140 x 80 см
    sentences:
      - гибкое стекло dekorelle 140 x 80 см
      - стекломагнезитовый лист стандарт 2500 x 1220 x 8 mm
      - труба valfex полипропилен арmир. стекл. d 40 x 5.5 mm pn 20 4 m
  - source_sentence: трап горизонтальный 100 x 100 mm d50 mm с гидрозатвором
    sentences:
      - >-
        лоток водоотводный standartpark basic dn100 1000 x 156 x 120 mm пластик
        8020 m
      - >-
        трап ани пласт горизонтальный d 50 mm 100 x 100 mm не регулируемый
        гидрозатвор
      - саморез шуж swfs pn 4 x 12 1000 шт
pipeline_tag: sentence-similarity
library_name: sentence-transformers
metrics:
  - cosine_accuracy
model-index:
  - name: SentenceTransformer based on google/embeddinggemma-300m
    results:
      - task:
          type: triplet
          name: Triplet
        dataset:
          name: hard neg eval
          type: hard-neg-eval
        metrics:
          - type: cosine_accuracy
            value: 0.9975224137306213
            name: Cosine Accuracy

SentenceTransformer based on google/embeddinggemma-300m

This is a sentence-transformers model finetuned from google/embeddinggemma-300m. 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: google/embeddinggemma-300m
Maximum Sequence Length: 2048 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': 2048, 'do_lower_case': False, 'architecture': 'Gemma3TextModel'})
  (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})
  (2): Dense({'in_features': 768, 'out_features': 3072, 'bias': False, 'activation_function': 'torch.nn.modules.linear.Identity'})
  (3): Dense({'in_features': 3072, 'out_features': 768, 'bias': False, 'activation_function': 'torch.nn.modules.linear.Identity'})
  (4): 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("Maksim-KOS/embeddinggemma-300m-saturn-planet")
# Run inference
queries = [
    "\u0442\u0440\u0430\u043f \u0433\u043e\u0440\u0438\u0437\u043e\u043d\u0442\u0430\u043b\u044c\u043d\u044b\u0439 100 x 100 mm d50 mm \u0441 \u0433\u0438\u0434\u0440\u043e\u0437\u0430\u0442\u0432\u043e\u0440\u043e\u043c",
]
documents = [
    'трап ани пласт горизонтальный d 50 mm 100 x 100 mm не регулируемый гидрозатвор',
    'лоток водоотводный standartpark basic dn100 1000 x 156 x 120 mm пластик 8020 m',
    'саморез шуж swfs pn 4 x 12 1000 шт',
]
query_embeddings = model.encode_query(queries)
document_embeddings = model.encode_document(documents)
print(query_embeddings.shape, document_embeddings.shape)
# [1, 768] [3, 768]

# Get the similarity scores for the embeddings
similarities = model.similarity(query_embeddings, document_embeddings)
print(similarities)
# tensor([[ 0.7530,  0.1274, -0.0019]])

Evaluation

Metrics

Triplet

Dataset: hard-neg-eval
Evaluated with TripletEvaluator

Metric	Value
cosine_accuracy	0.9975

Training Details

Training Dataset

Unnamed Dataset

Size: 123,502 training samples
Columns: anchor, positive, and negative

Approximate statistics based on the first 1000 samples:

	anchor	positive	negative
type	string	string	string
details	min: 6 tokens mean: 21.44 tokens max: 56 tokens	min: 7 tokens mean: 25.22 tokens max: 48 tokens	min: 9 tokens mean: 26.08 tokens max: 50 tokens

Samples:

anchor	positive	negative
`ремонтный состав быстротвердеющий gerkules 5 кг`	`ремонтный состав gerkules gs 22 быстротвердеющий 5 кг`	`клей для плитки и керамогранита gerkules granit pro gm 245 25 кг`
`тяга подвеса 350 mm`	`тяга подвеса 350 mm`	`подвес прямой 200 mm knauf 72519`
`подложка вспененный полиэтилен 4 mm`	`подложка из вспененного полиэтилена 4 mm 1 п.m.`	`подложка из вспененного полиэтилена 5 mm 50 m 2`

Loss: MultipleNegativesRankingLoss with these parameters:

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

Evaluation Dataset

Unnamed Dataset

Size: 13,723 evaluation samples
Columns: anchor, positive, and negative

Approximate statistics based on the first 1000 samples:

	anchor	positive	negative
type	string	string	string
details	min: 6 tokens mean: 21.29 tokens max: 57 tokens	min: 7 tokens mean: 25.17 tokens max: 50 tokens	min: 7 tokens mean: 26.08 tokens max: 50 tokens

Samples:

anchor	positive	negative
`краска для стен и потолков морозостойкая 6.5 кг`	`краска для стен и потолков pufas decoself морозостойкая 6.5 кг`	`краска моющаяся pufas decoself морозостойкая нв 23.1 кг`
`розетка компьютерная rj45 1 гнездо скрытая`	`розетка компьютерная rj 45 с у atlasdesign atn001083 1 гнездо карбон`	`розетка компьютерная rj 45 с у atlasdesign atn001085 2 гнезда карбон`
`саморез кровельный 4.8 x 28 коричневый оцинкованный`	`саморез кровельный swfs premium ral 8017 zp 4.8 x 28 300 шт круглый конт р`	`саморез кровельный swfs premium ral 5002 zp 4.8 x 28 250 шт`

Loss: MultipleNegativesRankingLoss with these parameters:

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

Training Hyperparameters

Non-Default Hyperparameters

eval_strategy: steps
per_device_train_batch_size: 32
per_device_eval_batch_size: 32
gradient_accumulation_steps: 8
learning_rate: 1e-05
weight_decay: 0.01
num_train_epochs: 10
lr_scheduler_type: cosine
warmup_ratio: 0.1
load_best_model_at_end: True
optim: adamw_torch
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: 8
eval_accumulation_steps: None
torch_empty_cache_steps: None
learning_rate: 1e-05
weight_decay: 0.01
adam_beta1: 0.9
adam_beta2: 0.999
adam_epsilon: 1e-08
max_grad_norm: 1.0
num_train_epochs: 10
max_steps: -1
lr_scheduler_type: cosine
lr_scheduler_kwargs: None
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
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: True
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
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: 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: 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	Validation Loss	hard-neg-eval_cosine_accuracy
0.1036	50	0.1448	-	-
0.2073	100	0.0509	0.0448	0.9884
0.3109	150	0.0373	-	-
0.4145	200	0.0285	0.0288	0.9923
0.5181	250	0.0291	-	-
0.6218	300	0.0264	0.0228	0.9944
0.7254	350	0.0252	-	-
0.8290	400	0.0236	0.0212	0.9953
0.9326	450	0.0258	-	-
1.0352	500	0.0217	0.0201	0.9956
1.1389	550	0.02	-	-
1.2425	600	0.0161	0.0182	0.9953
1.3461	650	0.0189	-	-
1.4497	700	0.0162	0.0164	0.9959
1.5534	750	0.0174	-	-
1.6570	800	0.0173	0.0160	0.9956
1.7606	850	0.0168	-	-
1.8642	900	0.015	0.0138	0.9972
1.9679	950	0.0132	-	-
2.0705	1000	0.0117	0.0129	0.9967
2.1741	1050	0.0111	-	-
2.2777	1100	0.0092	0.0120	0.9970
2.3813	1150	0.0117	-	-
2.485	1200	0.0101	0.0113	0.9972
2.5886	1250	0.0112	-	-
2.6922	1300	0.0098	0.0116	0.9972
2.7959	1350	0.0121	-	-
2.8995	1400	0.0095	0.0114	0.9975

The bold row denotes the saved checkpoint.

Framework Versions

Python: 3.12.3
Sentence Transformers: 5.2.0
Transformers: 4.57.6
PyTorch: 2.9.1+cu128
Accelerate: 1.12.0
Datasets: 4.5.0
Tokenizers: 0.22.2

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}
}