This is a sentence-transformers model finetuned from Snowflake/snowflake-arctic-embed-l-v2.0. It maps sentences & paragraphs to a 1024-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: Snowflake/snowflake-arctic-embed-l-v2.0
Maximum Sequence Length: 8192 tokens
Output Dimensionality: 1024 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: XLMRobertaModel 
  (1): Pooling({'word_embedding_dimension': 1024, '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})
  (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("haophancs/snowflake-arctic-embed-l-v2.0-pits")
# Run inference
sentences = [
    'Fixed-point observation test.\nPre-check.\nAll members.\n02-08-19_Steering_Switch.\nExternal SW.\nSteering SW( ).\nSOURCE DOWN.\n02-08-19-042.\n1_A.\nExternal SW>Steering SW( )>SOURCE DOWN',
    'Updating AM/FM broadcast station list ACC disconnection → Manual update of station list results in failure to detect broadcast stations.\nAudio.\nAM_FM.\nIllegal Behavior/Incorrect Action.\nRadio.\nST.\nAudioPF.\nAudioPF.\nWhen executing ACC OFF, if the previous source was FM/AM, issue a radio switch request (OFF) and then execute a radio switch request (FM or AM). Before the completion of the audio OFF power switch, the above-mentioned wireless power switch was requested..\nVehicle: Desktop Environment Actual Machine: Display: RD_AM Screen Audio: AM The radio preset mode should be "Broadcast List" The received electric field of the radio should have receivable broadcast stations in the Broadcast List Connected Devices: RADIO ANT Vehicle Signals: ACC ON, PKB ON, SPD OFF (0km), REV OFF, TAIL OFF.\n① Press [List Update] → Manual update of the broadcast station list begins ※ ONS "Cannot operate during broadcast station search" is displayed ② During manual update of the broadcast station list, perform ACC momentary disconnection ※ ACC momentary disconnection involves turning OFF/ON within less than 1 second ※ When ACC momentarily disconnects, the backlight will turn off for a moment but quickly returns to displaying ONS "Cannot operate during broadcast station search" ③ No operation until completion of broadcast station list update → ONS "No receivable broadcast stations TS[OK]" is displayed ★NG occurs ④ Press TS[OK] → ONS disappears ⑤ Press [List Update] → ONS "No receivable broadcast stations TS[OK]" is displayed ★NG occurs.\nDuring the ACC disconnection performed while updating the radio station list, the receivable radio stations will not be extracted during manual update of the station list, and ONS "No available broadcast stations" will be displayed. If you perform a manual update of the station list again after displaying ONS "No available broadcast stations", ONS "No available broadcast stations" will be displayed immediately after the manual update operation, making it impossible to manually update the station list..\nWhen manually updating the broadcast station list, search for receivable broadcast stations and extract up to 12 stations in order of reception sensitivity as the "Broadcast Station List.".\n2-1.No Use Case/Insufficient Use Cases.\nConsideration遗漏 in ACC instantaneous disconnection. It has entered a path where source switching is not executed after radio switch request (OFF).',
    "[Navi]After Heat Run, the navigation app crashes due to inactivity..\nNavi.\nNavigation common.\nApp crash.\nNavi.\nNaviApp.\nClear the list (vector) of icon rectangle information for collision detection at the timing when updating the FBO for rendering..\nDisplay: AM Audio: AM • Connected to rear seat • TV antenna connected • Radio antenna connected • Route setting: Available (with driving) • Bluetooth connection available • Wi-Fi connection available • Stack more than 7 activities and NaviApp's Activity (name search) is older than 7th position • ACC on.\nNo operation.\nAfter heat run, the navigation app crashes due to inactivity. ■Supplement · Although it does not recover with ACC off/on as shown in the video, we have confirmed that it recovers after turning ACC on about 10 minutes after ACC off. · The heat run operation time is over 80 hours..\nThe nav app does not crash even when left idle after a heat run..\n5-5. Remains of Past Design/Left as It Was from Past Design.\nAt the time of drawing execution, I am creating a list (vector array) of icon rectangle information for collision detection. However, this list is replaced with a new one after display execution. Therefore, in cases where display completion does not occur due to drawing cancellation or similar issues, the list continues to be registered without being cleared, resulting in unexpected memory allocation continuing indefinitely.",
]
embeddings = model.encode(sentences)
print(embeddings.shape)
# [3, 1024]

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

Evaluation

Metrics

Information Retrieval

Dataset: dim_1024
Evaluated with InformationRetrievalEvaluator

Metric	Value
cosine_accuracy@30	0.3063
cosine_accuracy@100	0.4505
cosine_accuracy@300	0.5946
cosine_precision@30	0.0102
cosine_precision@100	0.0045
cosine_precision@300	0.002
cosine_recall@30	0.3063
cosine_recall@100	0.4505
cosine_recall@300	0.5946
cosine_ndcg@10	0.0735
cosine_mrr@10	0.0513
cosine_map@100	0.0627

Training Details

Training Dataset

Unnamed Dataset

Size: 5,945 training samples
Columns: positive and anchor
Approximate statistics based on the first 1000 samples:
positive anchor
type string string
details
min: 31 tokens
mean: 167.44 tokens
max: 805 tokens
min: 183 tokens
mean: 366.94 tokens
max: 634 tokens

Samples:

positive	anchor
Fixed-point observation test.[object Object]Pre-check.[object Object]All members.[object Object]02-06-51_Drive Recorder Integration.[object Object]Drive recorder integration.[object Object]Collaboration function.[object Object]Image quality adjustment.[object Object]02-06-51-120.[object Object]2_B.[object Object]Drive recorder connection > Connection function > Image quality adjustment.[object Object]Screen: Drive Recorder File Playback (Map) Screen.[object Object]①Launch the picture adjustment screen and change the brightness and contrast settings..[object Object]・Adjustable brightness and contrast	[ ] When manually switching the camera connection determination during the connection of a wide camera (YOP), the guide lines for the rear camera's top-down view change..[object Object]Diag.[object Object]Camera_Diag.[object Object]Illegal Behavior/Incorrect Action.[object Object]camera.[object Object]ST.[object Object]VideoPF.[object Object]VideoPF.[object Object]Modify the possible range for setting the guideline vertex coordinates to include the footer area (0-799, 0-479). This will prevent the coordinate range check from considering it abnormal and ensure that the initial guide lines are not displayed..[object Object]・System Information Settings ・Wide Camera (YOP) Connected ・Vehicle Speed 0 km/h ・Parking Brake: ON.[object Object]① REV ON: Current guidelines are confirmed ② REV OFF ③ Transition to developer diag ④ Press TSW[Camera Settings] ⑤ Press TSW[Manual Switch for Camera Connection Judgment] ⑥ Press TSW[Wide Camera(YOP)]→Press [Yes] ⑦ ACC OFF, then ACC ON for more than 10 seconds ⑧ REV ON: Current guidelines are confirmed →★The guideline of the rear camera top-down view is changed.[object Object]When manually switching the came...
`Fixed-point observation test.[object Object]Pre-check.[object Object]All members.[object Object]02-50-54_Developer_Diag_Function_for_.[object Object]-1600.[object Object]3_C.[object Object]>>`	[ ] When manually switching the camera connection determination during the connection of a wide camera (YOP), the guide lines for the rear camera's top-down view change..[object Object]Diag.[object Object]Camera_Diag.[object Object]Illegal Behavior/Incorrect Action.[object Object]camera.[object Object]ST.[object Object]VideoPF.[object Object]VideoPF.[object Object]Modify the possible range for setting the guideline vertex coordinates to include the footer area (0-799, 0-479). This will prevent the coordinate range check from considering it abnormal and ensure that the initial guide lines are not displayed..[object Object]・System Information Settings ・Wide Camera (YOP) Connected ・Vehicle Speed 0 km/h ・Parking Brake: ON.[object Object]① REV ON: Current guidelines are confirmed ② REV OFF ③ Transition to developer diag ④ Press TSW[Camera Settings] ⑤ Press TSW[Manual Switch for Camera Connection Judgment] ⑥ Press TSW[Wide Camera(YOP)]→Press [Yes] ⑦ ACC OFF, then ACC ON for more than 10 seconds ⑧ REV ON: Current guidelines are confirmed →★The guideline of the rear camera top-down view is changed.[object Object]When manually switching the came...
Fixed-point observation test.[object Object]Pre-check.[object Object]All members.[object Object]02-06-58_Camera.[object Object]Rear view monitor.[object Object]Rear camera (back view monitor).[object Object]Caution message display.[object Object]1_A.[object Object]Back view monitor > Rear camera (back view monitor) > Caution message display.[object Object]Screen: Current Location Screen Voice: Audio OFF Camera: Normal Camera Connected.[object Object]①REV ON⇒"Please visually confirm the surroundings of the vehicle.".[object Object]・OSD (visual warning messages can be displayed)	[ ] When manually switching the camera connection determination during the connection of a wide camera (YOP), the guide lines for the rear camera's top-down view change..[object Object]Diag.[object Object]Camera_Diag.[object Object]Illegal Behavior/Incorrect Action.[object Object]camera.[object Object]ST.[object Object]VideoPF.[object Object]VideoPF.[object Object]Modify the possible range for setting the guideline vertex coordinates to include the footer area (0-799, 0-479). This will prevent the coordinate range check from considering it abnormal and ensure that the initial guide lines are not displayed..[object Object]・System Information Settings ・Wide Camera (YOP) Connected ・Vehicle Speed 0 km/h ・Parking Brake: ON.[object Object]① REV ON: Current guidelines are confirmed ② REV OFF ③ Transition to developer diag ④ Press TSW[Camera Settings] ⑤ Press TSW[Manual Switch for Camera Connection Judgment] ⑥ Press TSW[Wide Camera(YOP)]→Press [Yes] ⑦ ACC OFF, then ACC ON for more than 10 seconds ⑧ REV ON: Current guidelines are confirmed →★The guideline of the rear camera top-down view is changed.[object Object]When manually switching the came...

Loss: MatryoshkaLoss with these parameters:
json
1{ 2 "loss": "MultipleNegativesRankingLoss", 3 "matryoshka_dims": [ 4 1024 5 ], 6 "matryoshka_weights": [ 7 1 8 ], 9 "n_dims_per_step": -1 10}

Training Hyperparameters

Non-Default Hyperparameters

eval_strategy: epoch
gradient_accumulation_steps: 16
learning_rate: 2e-05
num_train_epochs: 4
lr_scheduler_type: cosine
warmup_ratio: 0.1
bf16: True
tf32: True
load_best_model_at_end: True
optim: adamw_torch_fused
batch_sampler: no_duplicates

All Hyperparameters

Click to expand

overwrite_output_dir: False
do_predict: False
eval_strategy: epoch
prediction_loss_only: True
per_device_train_batch_size: 8
per_device_eval_batch_size: 8
per_gpu_train_batch_size: None
per_gpu_eval_batch_size: None
gradient_accumulation_steps: 16
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: 4
max_steps: -1
lr_scheduler_type: cosine
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: True
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}
deepspeed: None
label_smoothing_factor: 0.0
optim: adamw_torch_fused
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
eval_use_gather_object: False
prompts: None
batch_sampler: no_duplicates
multi_dataset_batch_sampler: proportional

Training Logs

Epoch	Step	Training Loss	dim_1024_cosine_ndcg@10
0	0	-	0.0776
0.2151	10	1.6503	-
0.4301	20	1.4755	-
0.6452	30	1.3519	-
0.8602	40	1.478	-
1.0323	48	-	0.0840
1.0228	50	1.117	-
1.2379	60	1.3709	-
1.4530	70	1.2806	-
1.6680	80	1.1826	-
1.8831	90	1.2987	-
2.0336	97	-	0.0792
2.0457	100	0.9801	-
2.2608	110	1.2541	-
2.4758	120	1.1502	-
2.6909	130	1.1076	-
2.9059	140	1.1814	-
3.0349	146	-	0.0747
3.0685	150	0.8835	-
3.2836	160	1.1296	-
3.4987	170	1.0703	-
3.7137	180	1.013	-
3.7997	184	-	0.0735

The bold row denotes the saved checkpoint.

Framework Versions

Python: 3.11.11
Sentence Transformers: 3.3.1
Transformers: 4.44.2
PyTorch: 2.5.1+cu124
Accelerate: 1.3.0
Datasets: 2.19.0
Tokenizers: 0.19.1

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",
}

MatryoshkaLoss

bibtex

@misc{kusupati2024matryoshka,
    title={Matryoshka Representation Learning},
    author={Aditya Kusupati and Gantavya Bhatt and Aniket Rege and Matthew Wallingford and Aditya Sinha and Vivek Ramanujan and William Howard-Snyder and Kaifeng Chen and Sham Kakade and Prateek Jain and Ali Farhadi},
    year={2024},
    eprint={2205.13147},
    archivePrefix={arXiv},
    primaryClass={cs.LG}
}

MultipleNegativesRankingLoss

bibtex

@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}
}

	positive	anchor
type	string	string
details	min: 31 tokens mean: 167.44 tokens max: 805 tokens	min: 183 tokens mean: 366.94 tokens max: 634 tokens