petkopetkov/e5-large-v2-patent

SentenceTransformer based on intfloat/e5-large-v2

This is a sentence-transformers model finetuned from intfloat/e5-large-v2. 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: intfloat/e5-large-v2
• Maximum Sequence Length: 512 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': 512, 'do_lower_case': False}) with Transformer model: BertModel 
  (1): Pooling({'word_embedding_dimension': 1024, '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): 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("petkopetkov/e5-large-v2-patent")
# Run inference
sentences = [
    'query: LOCAL DECISION MAKING The present disclosure relates to the use of cryptographic techniques to facilitate local decision making at a gateway device interfacing between an operator device and edge devices, for example as can be found in Internet of Things infrastructures. Local decision making is facilitated in the context of end to end encryption of data between the edge device and operator device by enabling a function of the data to be computed without decrypting the data, for example using Functional Encryption (FE). Examples of edge devices are video surveillance cameras or utility consumption meters but the present disclosure is applicable to any other kind of edge device that produces data to be transmitted with end to end encryption. The disclosure is also not limited to IoT infrastructures. A method of local decision making in a communications system comprising an operator device in communication with a plurality of gateway devices and a plurality of edge devices in communication with each gateway device, the method comprising: receiving encrypted data at a gateway device from an edge device for transmission to the operator device, wherein the encrypted data is encrypted with a public key such that it can be decrypted with a corresponding private key and a function of the data can be computed without the private key, wherein the gateway device does not have access to the private key; applying, at the gateway device, an operator to the encrypted data to compute a value of a function of the input data without decrypting the encrypted data; determining, at the gateway device, an action to be taken based on the value; and taking the determined action. A method according to claim 1, wherein determining the action comprises deciding whether or how to transmit the encrypted input data to the operator and taking the determined action comprises transmitting or not transmitting the encrypted input data to the operator device based on the determination. A method according to claim 1 or 2 comprising receiving instances of encrypted data from respective edge devices and computing a respective value of the function for each instance of encrypted data, wherein determining the action comprises selecting which instances of the encrypted data to transmit and, or how to transmit them based on the computed values and the action comprises transmitting selected none, one or more of the received instances of encrypted data based on the computed values. A method according to claim 1, wherein the action comprises sending a control signal to the edge device and/or sending an alarm signal to the operator device. A method according to claim 1 or 4, comprising receiving instances of encrypted data from respective edge devices and computing a respective value of the function for each instance of encrypted data and determining the action to be taken based on the computed values, wherein the action comprises sending a control signal to the edge device and/or sending an alarm signal to the operator device. --> A method according to claim 1, 2 or 3, wherein the edge device or devices comprise a video camera, the encrypted input data comprises encrypted video data and encrypted motion data indicative of motion in frames of the video data and the function comprises a function of the motion data indicative of the amount of motion in frames of the video. A method according to claim 6, wherein the action comprises selecting a quality of video data to be transmitted to the operator based on the value. A method according to claim 1, 4 or 5, wherein the edge device or devices comprise a utility consumption meter and the encrypted input data comprises utility consumption data. A method according to any preceding claim, wherein the function comprises a sum. A method according to any preceding claim, wherein the function returns a value indicating whether a sum of magnitudes of the data exceed a threshold value. A method according any preceding claim, wherein the input data is encrypted using functional encryption. A gateway device for use in a communications system comprising an operator device in communication with a plurality of gateway devices and a plurality of edge devices in communication with each gateway device, wherein the gateway device is configured to implement a method according to any preceding claim and comprisesa memory storing computer instructions that, when run on a processor implement the method;a communications interface for receiving the encrypted input data from an edge device and for transmitting the encrypted data to an operator device; anda processor configured to execute the computer instructions to implement the method. A communications system comprising an operator device in communication with a plurality of gateway devices according to claim 11 and a plurality of edge devices in communication with each gateway device. A communications system according to claim 12, wherein connections between the gateway and edge devices have higher bandwidth and/or shorter latency than connections between the edge devices and the operator device.',
    'passage: INFORMATION AND COMMUNICATION PROCESSING SYSTEM, METHOD, AND NETWORK NODE In a distributed information communication processing system in which a plurality of information communication devices provides a service through a network, response speed, electric power consumption, and further reliability are improved. The distributed information communication processing system which provides various services is configured by associating an entrance node (EN) which executes filtering near sensors, actuators, and terminals being information sources, an intelligent node (IN) which changes an information processing position and executes information processing and communication processing instead of a data center (DC), and a management node (MN) which manages these nodes. Thereby, real time type information processing can be realized. An information communication processing system in which an information processing device which can execute an application and a plurality of terminals requesting services are located, the system comprising: a first network node connected to the terminals; a second network node connected to the first network node though a first network and connected to the information processing device through a second network; and a management node for managing the first network node and the second network node, wherein the first network node sends a packet which has a destination to the information processing device through the second network node; the second network note analyzes the packet when the packet is received, outputs the packet to a processor when the second network node has the processor which can perform processes associated with the analyzed result, the packet, and the application, and sends a packet including the processed result by the processor to the first network node; and the information processing device executes the application when the packet is received through the second network node. The information communication processing system according to claim 1,wherein the first network node comprises: an interface which sends and receives a packet; and a processing unit which processes the packet which the interface receives, and --> wherein the processing unit performs calculation processing, filtering, or aggregation processing to the packet received from the terminals, further selects the processed result, and sends the selected result. The information communication processing system according to claim 2,wherein the processing unit generates a processing command based on the packet received from the second network node, and sends the processing command to the terminals. The information communication processing system according to claim 2,wherein, the processing unit sends stored data stored in the first network node through the interface when the received packet is a stored data request. The information communication processing system according to claim 2,wherein the management node comprises a management table which stores contents of the calculation processing, the filtering, or the aggregation processing of the first node, and sends the contents of the management table to the first network node. The information communication processing system according to claim 1,wherein the second network node comprises: a plurality of processors, the processors executing any application; and --> a communication control unit which can transfer the received packet to at least any one of destinations among the processors and external nodes other than the second network node. The information communication processing system according to claim 6,wherein the communication control unit associates the packets which correspond with at least a part of a header and a payload of the packets with the same flow according to a predetermined rule, comprises a table which indicates a destination of the flow, and changes a destination of the packets to the destination of the associated flow according to the table. The information communication processing system according to claim 7,wherein the communication control unit rewrites the destination in the table based on load information of the processors. The information communication processing system according to claim 7,wherein the management node generates a request for replicating the application which is performed in the information processing device or a request for rewriting the destination in the table to the second network node, and sends to the second network node; andwherein the second network node replicates the application or processing associated with the application in --> the processor or rewrite the destination in the table based on the rewriting request according to the request from the management node. The information communication processing system according to claim 6,wherein the management node comprises a management table which stores nominated information for changing an execution target of the application or the processing associated with the application which is executed by the processors of the second network node, and sends contents of the management table to the second network table. The information communication processing system according to claim 1,wherein processing associated with the application executed on the second network node is processing in which a processing result is notified to the first network node early, compared with an application executed on the information processing device. An information communication processing method for providing a service to a plurality of terminals, the method comprising the steps of: connecting a first network node connected to the terminal to a second network node through a first network; connecting the second network node to an information processing device through a second network; and for providing the service to the terminals, sending a packet which has a destination to the --> information processing device to the first network by using information obtained from the terminal in the first network node; outputting the packet to an information processing function unit which the second network node has or a node other than the second network node based on a destination and information included in the packet by the second network node when second network node receives the packet from the first network; sending a packet including a processing result to the packet processed by the information processing function unit to the first network node by the second network node; and receiving the packets including the processed result by the first network node to provide the service to the terminal. A second network node sending and receiving a packet through a first network connected to a first network node to which a plurality of terminals are connected and a second network connected to an information processing, the second network node comprising: a network interface unit connected to the first network and the second network; a communication control unit analyzing the packet received through the network interface unit and transferring the packet to any destination; and an information processing function unit to which the packet received through the network interface unit are transferred by the communication control unit and which executes a predetermined application for the packets. The second network node according to claim 13,wherein, by using a table which records a connection status and a destination made of the packet, the communication control unit changes a destination of the packets based on the destinations in the table. The second network node according to claim 14,wherein the communication control unit rewrites the destination of the flow based on load information of the information processing function unit. The second network node according to claim 14,wherein the communication control unit rewrites the destination of the flow in which the connection status is an unconnected status based on a request from the management node. The second network node according to claim 14,wherein the communication control unit rewrites the destination of the flow to the information processing function unit. The second network node according to claim 14,wherein the information processing function unit comprises a plurality of processors, and the communication control unit changes a destination of the packet to any one of the processors. The information communication processing method according to claim 12,wherein the terminals comprise a monitoring camera and --> an automatic door;the first network node extracts a face by the monitoring camera and sends extracted face image data to the second network node; andthe second network node sends a control signal for opening the automatic door to the first network node when the face image data corresponds to a face image database. The information communication processing method according to claim 12,wherein the terminals comprise a sensor and a monitoring camera;the first network node sends output of the sensor and image data of the monitoring camera to the second network node when the sensor output of the sensor exceeds a set threshold value; andthe second network node sends the image data to a previously registered user when the second network node detects an abnormal value from the sensor output. The information communication processing method according to claim 12,wherein the terminals comprise an acceleration sensor or a vibration sensor;the first network node separates effective quake data and ineffective quake data based on an output of the acceleration sensor or the vibration sensor, and sends the separated effective quake data to the second network node;, andthe second network node generates an alarm notification based on the received effective quake data, and sends the --> generated alarm notification to a previously registered user. The information communication processing method according to claim 12,wherein the terminals comprise an sensor and a camera; the first network node stores a sensor output of the sensor and image data of the camera, generates corresponding event information when the output of the sensor exceeds a set threshold value, and sends the event information to the second network node; andthe second network node sends a transfer request of the necessary sensor output and the necessary image data to the first network node based on the event information.',
    'passage: INTRODUCER SHEATH, PLACEMENT DEVICE FOR BLOOD VESSEL TREATMENT INSTRUMENT, AND METHOD FOR SHORTENING INTRODUCER SHEATH A placement device (10) for a blood vessel treatment instrument has an introducer sheath (14) functioning as an outer tube, and also has an inner tube (16). The introducer sheath (14) has a flexible tube-shaped sheath body (18) and a hub (20) into which the base end of the sheath body (18) is inserted. The hub (20) takes up the base end of the sheath body (18) into the hub (20) by means of take-up shafts (30, 32) while tearing the base end of the sheath body (18) by cutting blades (72, 74), and thus the length of extension of the sheath body (18) from the hub (20) can be shortened. An introducer sheath (14, 14a, 14b) into which a long shaft is inserted, comprising: a flexible tube-shaped sheath body (18); and a hub (20) into which a proximal portion of the sheath body (18) is inserted, wherein the hub (20) takes up a proximal portion of the sheath body (18) into the hub (20) while tearing the proximal portion of the sheath body (18), whereby the length of extension of the sheath body (18) from the hub (20) can be shortened. The introducer sheath (14, 14a, 14b) according to claim 1,wherein the hub (20) has: a cutting section (28, 120) by which slits along an axial direction are formed in the sheath body (18); and a plurality of take-up shafts (30, 32) for respectively taking up terminal pieces of the sheath body (18) torn by the slits. The introducer sheath (14, 14a, 14b) according to claim 2, -->wherein the cutting section (28, 120) forms the slits in portions of the sheath body (18) which are on opposite sides with respect to a circumferential direction, andthe plurality of take-up shafts (30, 32) are two take-up shafts (30, 32) disposed at positions spaced from each other along the direction of splitting of the sheath body (18) by the slits. The introducer sheath (14, 14a) according to claim 2,wherein the cutting section (28, 120) has a plurality of cutting blades (72, 74) by which the slits are formed in circumferential-directionally different portions of the sheath body (18). The introducer sheath (14, 14a, 14b) according to claim 2,wherein the hub (20) further has an interlocking mechanism (106) by which the plurality of take-up shafts (30, 32) are rotated in an interlocked manner. The introducer sheath (14a) according to claim 2,wherein the hub (20) has rotation restraining --> mechanisms (104, 105) for inhibiting the take-up shafts (30, 32) from rotating in an unwinding direction. The introducer sheath (14, 14a, 14b) according to claim 2,wherein the hub (20) has: a hub body (31) provided with a hollow section in which the take-up shafts (30, 32) and the proximal portion of the sheath body (18) are housed; and a rotational operating section (34) which is rotationally operated from outside of the hub body (31) to thereby rotate the take-up shafts (30, 32), the hub body (31) being configured to be liquid-tight so that a liquid flowing into the inside of the hub (20) through the sheath body (18) does not leak to the exterior. A blood vessel treatment instrument placement device (10) by which a blood vessel treatment instrument (12) having a self-expanding function is fed to and placed indwelling in a desired treatment site in a blood vessel,wherein the placement device (10) includes an introducer sheath (14, 14a, 14b) having a sheath --> body (18) for housing the blood vessel treatment instrument (12) on an inner circumference of a distal portion thereof, andan inner tube (16) slidably inserted inside the sheath body (18);the introducer sheath (14, 14a, 14b) hasthe sheath body (18) which is flexible and tube-like in shape, anda hub (20) in which a proximal portion of the sheath body (18) is inserted; andthe hub (20) takes up a proximal portion of the sheath body (18) into the hub (20) while tearing the proximal portion of the sheath body (18), whereby the length of extension of the sheath body (18) from the hub (20) can be shortened. A method for shortening an introducer sheath (14, 14a, 14b) in which a long shaft is inserted, the method comprising: tearing a proximal portion of a sheath body (18); drawing the sheath body (18) into the hub (20) in which the proximal portion of the sheath body (18) is inserted; and taking up terminal pieces of the sheath body (18) --> having been torn.',
]
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: patent_dev
• Evaluated with InformationRetrievalEvaluator

Metric	Value
cos_sim_accuracy@1	0.4751
cos_sim_accuracy@3	0.6754
cos_sim_accuracy@5	0.7442
cos_sim_accuracy@10	0.8406
cos_sim_precision@1	0.4751
cos_sim_precision@3	0.2554
cos_sim_precision@5	0.1746
cos_sim_precision@10	0.1018
cos_sim_recall@1	0.4118
cos_sim_recall@3	0.6289
cos_sim_recall@5	0.7006
cos_sim_recall@10	0.8075
cos_sim_ndcg@10	0.6323
cos_sim_mrr@10	0.5941
cos_sim_map@100	0.5749

Training Details

Training Dataset

Unnamed Dataset

• Size: 7,726 training samples
• Columns: sentence_0 and sentence_1
• Approximate statistics based on the first 1000 samples:

  	sentence_0	sentence_1
  type	string	string
  details	min: 306 tokens mean: 510.19 tokens max: 512 tokens	min: 312 tokens mean: 508.81 tokens max: 512 tokens

• Samples:

  sentence_0	sentence_1
  query: TISSUE RESECTING INSTRUMENT INCLUDING AN OUTFLOW CONTROL SEAL A tissue resecting instrument (10) includes an end effector assembly (100) having a proximal hub housing (110), outer shaft (120) and inner shaft (130) extending therefrom, and an inner core drive assembly to rotate and reciprocate the inner shaft relative to the outer shaft. The inner core drive assembly includes a proximal receiver that receives a rotational input and rotates in response and includes a seal member disposed thereon. The rotation of the proximal receiver effects rotation of a connector and reciprocation of the connector between a proximal position and a distal position. The connector is operably coupled to the inner shaft such that the rotation and reciprocation of the connector effects the rotation and reciprocation of the inner shaft. In the proximal position, the connector and the seal member establish a seal that blocks outflow. In the distal position, the connector is displaced from the seal memb...	passage: Reciprocating rotary arthroscopic surgical instrument A surgical instrument includes a cutting member with an implement for cutting tissue, and a drive coupled to the cutting member. The drive may include a drive member having a helical groove and being attached to the cutting member. Furthermore, the drive may include an inner drive hub coupled to the drive member such that the drive member rotates with the inner drive hub while being free to translate relative to the inner drive hub. The drive simultaneously rotates and translates the cutting member in response to a force applied to the drive. A surgical instrument (400, 600), comprising: a cutting member (185, 285) including an implement (182, 282) for cutting tissue; and a drive (110) coupled to the cutting member (185, 285) to simultaneously rotate and translate the cutting member (185, 285) in response to a force applied to the drive,wherein the drive (110) includes a drive member (450, 650) coupled to the cutting member...
  query: VEHICLE LAMP CONTROL SYSTEM A vehicle lamp system 1 includes a light source unit (10) capable of individually adjusting an illuminance of light to be radiated to each of a plurality of individual areas ahead of a host vehicle; an imaging unit (12) configured to take an image ahead of the host vehicle; a high-speed low-accuracy analysis unit (114) configured to detect luminance of each individual area based on information obtained from the imaging unit; a low-speed high-accuracy analysis unit (116) configured to detect target objects ahead of the host vehicle based on the information obtained from the imaging unit; a tracking unit (40) configured to determine a specific target object from the target objects detected by the low-speed high-accuracy analysis unit and to detect displacement of the specific target object based on a detection result of the high-speed low-accuracy analysis unit; an illuminance setting unit (42) configured to set, based on the detection result of the hig...	passage: HEADLAMP CONTROLLER In a control system (14), an OF calculating means (58) calculates the OF of an object existing in front of a vehicle as a light emitter or a light reflector from the brightness information of an acquired pick-up image in front of the vehicle. An object attribute determining means (60) determines the attribute of the object according to the OF. A light distribution control ECU (34) controls the light distribution of headlamp units (12R, 12L) provided in the vehicle according the attribute of the object. An image analysis means (50) estimates the shape of the road in front of the vehicle. The object attribute determining means (60) determines the attribute of the object according to the OF of the object and the shape of the road. A headlamp controller comprising: an optical flow calculation means configured to calculate an optical flow of an object present in front of a vehicle as a light emitter or light reflector, based on the luminance information of an ac...
  query: PRE DIFFUSER FOR A GAS TURBINE ENGINE A pre-diffuser (100) for a gas turbine engine (20) includes an exit guide vane ring (104) having a multiple of exit guide vanes (108) defined around an engine longitudinal axis (A), a hot fairing structure (102) adjacent to the exit guide vane ring (10) to define a multiple of diffusion passages (120) around the engine longitudinal axis (A), an outer radial interface (190) between a radial outer surface of the hot fairing structure (102) and the exit guide vane ring (104), the outer radial interface (190) being a full hoop structure, and an anti-rotation feature (130) between the hot fairing structure (120) and the exit guide vane ring (104), the anti-rotation feature (130) inboard of the multiple of diffusion passages (120). A pre-diffuser (100) for a gas turbine engine (20), comprising: an exit guide vane ring (104) having a multiple of exit guide vanes (108); a hot fairing structure (102) adjacent to the exit guide vane ring (104) to form...	passage: HIGH COMPRESSOR EXIT GUIDE VANE ASSEMBLY TO PRE-DIFFUSER JUNCTION A pre-diffuser and exit guide vane (EGV) system for a gas turbine engine (10) includes an annular EGV assembly (31) containing a number of guide vanes (33) and having an annular opening bounded by a radially inner annular sealing surface at a first radius and a radially outer annular sealing surface at a second radius. First and second seals (35) substantially matching the first and second radii respectively join the EGV assembly (31) to an annular pre-diffuser (30) having an annular opening bounded by radially inner and outer annular sealing surfaces at substantially the first and second radii. The seals (35) seal the inner sealing surface of the EGV assembly (31) to the inner sealing surface of the pre-diffuser (30) and the second seal (35) seals the outer sealing surface of the EGV assembly (31) to the outer sealing surface of the pre-diffuser (30), such that the EGV assembly annular opening is in fluid commu...

• Loss: MultipleNegativesRankingLoss with these parameters:

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

Training Hyperparameters

Non-Default Hyperparameters

• eval_strategy: steps
• 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: 8
• per_device_eval_batch_size: 8
• 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: 3
• 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}
• tp_size: 0
• 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
0.5176	500	0.1257
1.0	966	-
1.0352	1000	0.0747
1.5528	1500	0.0405
2.0	1932	-
2.0704	2000	0.031
2.5880	2500	0.0198

Framework Versions

• Python: 3.12.3
• Sentence Transformers: 4.0.2
• Transformers: 4.51.2
• PyTorch: 2.6.0+cu124
• Accelerate: 1.6.0
• Datasets: 3.5.0
• Tokenizers: 0.21.1

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}
}