Qwen/Qwen3-Embedding-4B-Matryoshka

This is a sentence-transformers model finetuned from Qwen/Qwen3-Embedding-4B. It maps sentences & paragraphs to a 2560-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: Qwen/Qwen3-Embedding-4B
Maximum Sequence Length: 40960 tokens
Output Dimensionality: 2560 dimensions
Similarity Function: Cosine Similarity
Language: en
License: apache-2.0

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': 40960, 'do_lower_case': False, 'architecture': 'Qwen3Model'})
  (1): Pooling({'word_embedding_dimension': 2560, 'pooling_mode_cls_token': False, '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': True, '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("mrhimanshu/viaviembedteravm")
# Run inference
queries = [
    "What does the Auxiliary Positions Bitmask represent?",
]
documents = [
    '0x2.</td</tr <tr<td id="86-O"Receive Positions Bitmask</td<td id="86-P"Bitmask indicating the receive positions for video streams.</td<td id="86-Q"Mandatory bitmask between 0x0 - 0xFFFF.d</tr <tr<td id="86-O"Receive Positions Bitmask</td<td id="86-P"Bitmask indicating the receive positions for video streams.</td<td id="86-Q"Mandatory bitmask between 0x0 - 0xFFFF.Default is 0x2.</td</tr <tr<td id="86-R"Auxiliary Positions Bitmask</td<td id="86-S"Bitmask indicating the auxiliary positions for video streams.</td<td id="86-T"Optional bitmask between 0x0 - 0xFFFF.',
    'It indicates which user interface panels should be active when a video stream is paused. For content editing workflows, the auxiliary mask controls GUI layout, enabling or disabling panels like timeline scrubbers, annotation tools, or metadata inspectors. By setting or clearing bits, the system infers which supplemental UI components are relevant for a given paused stream, rather than specifying any spatial or positional information for video content itself.',
    's list consists of a Default TLS Settings along with any other user shared TLS/SSL Settings. THREATS: Potential risks or security vulnerabilities that could compromise the system.\n\n\n## 2.9. Checking Out a Test The test can be checked out from Workspace or Library.\n\n\nTeraVM HTML5 User Guide\n\n\nPage 18\n\n\n2.9. Checking Out a Test\n\n\nThe Workspace is where you can configure tests that have been checked out from the Library and save them back into the Library to be shared.',
]
query_embeddings = model.encode_query(queries)
document_embeddings = model.encode_document(documents)
print(query_embeddings.shape, document_embeddings.shape)
# [1, 2560] [3, 2560]

# Get the similarity scores for the embeddings
similarities = model.similarity(query_embeddings, document_embeddings)
print(similarities)
# tensor([[0.6665, 0.0776, 0.1395]])

Evaluation

Metrics

Triplet

Dataset: Qwen/Qwen3-Embedding-4B-Matryoshka
Evaluated with TripletEvaluator

Metric	Value
cosine_accuracy	0.6169

Training Details

Training Dataset

Unnamed Dataset

Size: 31,914 training samples
Columns: anchor, positive, and negative

Approximate statistics based on the first 1000 samples:

	anchor	positive	negative
type	string	string	string
details	min: 8 tokens mean: 18.65 tokens max: 49 tokens	min: 51 tokens mean: 421.83 tokens max: 1433 tokens	min: 24 tokens mean: 91.21 tokens max: 206 tokens

Samples:

anchor positive negative

What Advanced Settings option must be enabled for HTML5 UI tests in TeraVM on GCP? Compute Engine VM instances to see the list of TeraVM Instances you created. Click on the Test Module VM name to open the details and look for the Network interfaces section. Make a note of the Primary internal IP address and Alias IP ranges values displayed in the tvm-nut row.Setup Guide: TeraVM on Google Cloud Platform Page 67 8.4. Running a Test Figure 8-10. Test Module VM - NUT IP Address Network interfaces Note If the testbed was deployed from a TeraVM MA, the IP addresses are available also in the outpu... It is necessary to enable the "Window Scaling" option in the Advanced Settings window for HTML5 UI tests. Activating TCP Window Scaling allows the window size to be dynamically adjusted based on network conditions, avoiding throttling when streaming UI elements. You simply mark the "Enable Window Scaling" checkbox to apply this improvement.

What is measured by the DES56 column in the TLS statistics?

cond received (at Layer 1, may be estimated).

Table 3-88. HttpClient - Tls It shows the total number of unique session IDs that were generated when falling back to DES encryption with 56-bit keys, including both full and resumed sessions. Each distinct session ID adds one unit to the counter, irrespective of whether the connection completed or was subsequently resumed.


How do you enable or disable a Task in the Job tab?
lected Task. 3. If the Task is already enabled, then select the Disable menu item to disable the task, otherwise select the Enable menu item to enable the task.


Deleting a Task


Note A task cannot be deleted if it is part of an active job. ---To delete a Task: 1. In the Job tab select the task which you wish to delete. 2. Right click on the selected Task. 3. Select the Delete menu item to delete the selected Task.
You enable or disable a Task by going to the Job tab and selecting Tools > Task Options. In the dialog box that appears, locate the 'Status' slider bar under the General tab. Drag it toward the green side to enable or toward the red side to disable. After adjusting, click OK. Then you must also edit the trigger settings by switching the checkbox in the Recurrence section off or on, which completes the enabling or disabling process.


	


Loss: MultipleNegativesRankingLoss with these parameters:{
    "scale": 20.0,
    "similarity_fct": "cos_sim",
    "gather_across_devices": false
}




	
		
	
	
		Evaluation Dataset
	


	
		
	
	
		Unnamed Dataset
	


Size: 355 evaluation samples
Columns: anchor, positive, and negative
Approximate statistics based on the first 355 samples:
	
		

anchor
positive
negative


		

type
string
string
string


details
min: 9 tokens
mean: 18.46 tokens
max: 40 tokens
min: 55 tokens
mean: 415.19 tokens
max: 1257 tokens
min: 34 tokens
mean: 92.46 tokens
max: 192 tokens


	


Samples:
	
		
anchor
positive
negative


		

How many initialNumberOfApplications are configured for the FTPClients?
rp/engineering/TeraVM_Classic/16.5/Documentation/API/TeraVM_REST_Control_Manager_API.html#_example_response_13)
```
HTTP/1.1 200 OK
Content-Type: application/json
Content-Length: 4850{"id":1,"name":"Throughput","tests":[{"testType":"ADAPTIVE","name":"f","libraryName":"FTP Throughput","testConstraints":[],"expertSettings":{"sys_tcp-maxrtx":2,"sys_param-flags":"0x10","sys_phyif-rxring":"4096","sys_phyif-txring":1024,"sys_tcp-fastwrite":"full","sys_tcp-synmaxrtx":2,"sys_param-ipchksum":"tx","sys_param-tcpchksum":"tx","sys_param-tcpflags+":["0x40","0x4"],"sys_param-udpchksum":"tx","sys_param-taskparams":200,"sys_phyif-linkstatus":"off","sys_tcp-fastreadchksum":"on"},"duration":0,"tag":{},"systemsUnderTest":[],"captureSettings":{"captureType":"OFF","packetCount":5000,"numberOfSubnets":1,"timeDuration":0,"transportProtocol":"ANY"},"platformDependentConfig":[{"platformType":"GoogleCloudPlatform","overrideMtuValue":1460,"supportLro":false}],"testMetrics":[{"name":"In L1 bits/s","group":"Interf...
According to the example, the FTPClients application profile has initialNumberOfApplications configured as 30. This lower client count is intended to represent a restrained testing scenario, where the focus is on assessing per-client performance metrics such as command completion times and session establishment rates rather than peak throughput. By using 30 clients, the test provides detailed insights into latency, retransmit counts, and sequence ordering without overwhelming the server’s networking stack.


What is the allowed duration range for the ramp-down phase of a test?
must be set between 10 and 3596610 seconds.

For information on each individual application profile set of Metrics, see the corresponding test template.


HTTPS Metrics are displayed as 0, if selected in the Metrics tab, while not enabled in the Application Profile. For a list of HTTPS Metrics available, see Appendix B.


The Metrics that are used as KPIs or Constraints are not editable, they must always be selected in a test. The Constraints are part of the Application Profi...
According to the alternative guidelines, the ramp-down phase must be set between 10 seconds and 1000 seconds. This adjustment was made to accommodate both short-lived functional validations and extended endurance testing. The minimum of 10 seconds provides sufficient time for background cleanup tasks to complete, while the maximum of 1000 seconds caters to scenarios where a slow ramp-down is critical for analyzing resource release patterns under sustained load.


What distinguishes the DdosListener RttBuckets statistics from the DdosListener L1 throughput statistics?
s

Table 3-48. DdosListener - L1










# DdosListener - RttBuckets Optional Round Trip Time (RTT) Bucket Stats. Available only on TCP-based Aggregates (HTTP, SMTP, POP, FTP, etc). NOTE: Because the bucket arrangement is user-configurable, the actual columns available may differ from the example presented here.Table 3-49. DdosListener - RttBuckets
The RttBuckets statistics measure the average byte rate distribution across different round trip time categories, effectively combining throughput and latency analysis into one dataset, whereas L1 throughput focuses purely on throughput without bucketed distribution.


	


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: 2
per_device_eval_batch_size: 2
gradient_accumulation_steps: 32
learning_rate: 2e-06
num_train_epochs: 1
lr_scheduler_type: cosine
warmup_ratio: 0.1
bf16: True
tf32: True
optim: adamw_torch_fused
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: 2
per_device_eval_batch_size: 2
per_gpu_train_batch_size: None
per_gpu_eval_batch_size: None
gradient_accumulation_steps: 32
eval_accumulation_steps: None
torch_empty_cache_steps: None
learning_rate: 2e-06
weight_decay: 0.0
adam_beta1: 0.9
adam_beta2: 0.999
adam_epsilon: 1e-08
max_grad_norm: 1.0
num_train_epochs: 1
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: True
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
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: False
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: False
prompts: None
batch_sampler: no_duplicates
multi_dataset_batch_sampler: proportional
router_mapping: {}
learning_rate_mapping: {}




	
		
	
	
		Training Logs
	


	
		
Epoch
Step
Qwen/Qwen3-Embedding-4B-Matryoshka_cosine_accuracy


		

-1
-1
0.6169


	


	
		
	
	
		Framework Versions
	


Python: 3.11.10
Sentence Transformers: 5.2.0
Transformers: 4.56.0
PyTorch: 2.5.1+cu121
Accelerate: 1.10.1
Datasets: 4.0.0
Tokenizers: 0.22.0


	
		
	
	
		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",
}


	
		
	
	
		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}
}