Qwen3-Embedding-0.6B-18-layers

SentenceTransformer based on Qwen/Qwen3-Embedding-0.6B

This is a sentence-transformers model finetuned from Qwen/Qwen3-Embedding-0.6B on the nq dataset. 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.

[!NOTE] This is a first experiment attempting to distil the powerful Qwen/Qwen3-Embedding-0.6B model from 28 layers down to some lower layer count, in an attempt to speed up inference with minimal performance reductions.

To be specific, this early model falls from ~0.55 NDCG@10 average across NanoMSMARCO, NanoNFCorpus, and NanoNQ with the full 28 layers, to 0.5155 NDCG@10 on that selection with just 18 layers. Early tests indicate that using only 18 layers results in a 1.51x speedup compared to the full model.

This model was distilled using only 200k texts from one domain, reaching superior performance should be possible, especially with stronger distillation techniques like MarginMSE.

Model Details

Model Description

• Model Type: Sentence Transformer
• Base model: Qwen/Qwen3-Embedding-0.6B <!-- at revision b22da495047858cce924d27d76261e96be6febc0 -->
• Maximum Sequence Length: 512 tokens
• Output Dimensionality: 1024 dimensions
• Similarity Function: Cosine Similarity
• Training Dataset:
  • nq
• Language: en

<!-- - **License:** Unknown -->

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: Qwen3Model 
  (1): Pooling({'word_embedding_dimension': 1024, '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("tomaarsen/Qwen3-Embedding-0.6B-18-layers")
# Run inference
sentences = [
    'The actress was thirteen when she was offered the role of Annie.',
    'Contrasting significantly from other soccer leagues in the U.S., WLS intends to be an open entry, promotion and relegation competition.',
    'Narsingh Temple is situated at the across of the village just across confluence of Magri State village.',
]
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]

<!-- ### Direct Usage (Transformers) <details><summary>Click to see the direct usage in Transformers</summary> </details> --> <!-- ### Downstream Usage (Sentence Transformers) You can finetune this model on your own dataset. <details><summary>Click to expand</summary> </details> --> <!-- ### Out-of-Scope Use *List how the model may foreseeably be misused and address what users ought not to do with the model.* -->

Evaluation

Metrics

Information Retrieval

• Datasets: NanoMSMARCO, NanoNFCorpus and NanoNQ
• Evaluated with <code>InformationRetrievalEvaluator</code> with these parameters:

  {
      "query_prompt": "Instruct: Given a web search query, retrieve relevant passages that answer the query\nQuery:"
  }
  

Nano BEIR

• Dataset: NanoBEIR_mean
• Evaluated with <code>NanoBEIREvaluator</code> with these parameters:

  {
      "dataset_names": [
          "msmarco",
          "nfcorpus",
          "nq"
      ],
      "query_prompts": {
          "msmarco": "Instruct: Given a web search query, retrieve relevant passages that answer the query\nQuery:",
          "nfcorpus": "Instruct: Given a web search query, retrieve relevant passages that answer the query\nQuery:",
          "nq": "Instruct: Given a web search query, retrieve relevant passages that answer the query\nQuery:"
      }
  }
  

Knowledge Distillation

• Evaluated with <code>MSEEvaluator</code>

<!-- ## Bias, Risks and Limitations *What are the known or foreseeable issues stemming from this model? You could also flag here known failure cases or weaknesses of the model.* --> <!-- ### Recommendations *What are recommendations with respect to the foreseeable issues? For example, filtering explicit content.* -->

Training Details

Training Dataset

nq

• Dataset: nq at f9e894e
• Size: 197,462 training samples
• Columns: <code>text</code> and <code>label</code>
• Approximate statistics based on the first 1000 samples:

  	text	label
  type	string	list
  details	<ul><li>min: 27 tokens</li><li>mean: 89.38 tokens</li><li>max: 505 tokens</li></ul>	<ul><li>size: 1024 elements</li></ul>

• Samples:

  text	label
  <code>Instruct: Given a web search query, retrieve relevant passages that answer the query<br>Query:the movie bernie based on a true story</code>	<code>[-0.05126953125, -0.0020294189453125, 0.00152587890625, 0.060791015625, 0.022216796875, ...]</code>
  <code>College World Series The College World Series, or CWS, is an annual June baseball tournament held in Omaha, Nebraska. The CWS is the culmination of the National Collegiate Athletic Association (NCAA) Division I Baseball Championship tournament—featuring 64 teams in the first round—which determines the NCAA Division I college baseball champion. The eight participating teams are split into two, four-team, double-elimination brackets, with the winners of each bracket playing in a best-of-three championship series.</code>	<code>[0.033935546875, -0.0908203125, -0.010498046875, 0.0625, -0.01263427734375, ...]</code>
  <code>Instruct: Given a web search query, retrieve relevant passages that answer the query<br>Query:does the femoral nerve turn into the saphenous nerve</code>	<code>[0.052978515625, -0.0028228759765625, -0.0022430419921875, 0.0732421875, 0.044677734375, ...]</code>

• Loss: <code>MSELoss</code>

Evaluation Datasets

nq

• Dataset: nq at f9e894e
• Size: 3,000 evaluation samples
• Columns: <code>text</code> and <code>label</code>
• Approximate statistics based on the first 1000 samples:

  	text	label
  type	string	list
  details	<ul><li>min: 21 tokens</li><li>mean: 87.24 tokens</li><li>max: 410 tokens</li></ul>	<ul><li>size: 1024 elements</li></ul>

• Samples:

  text	label
  <code>Instruct: Given a web search query, retrieve relevant passages that answer the query<br>Query:who was the heir apparent of the austro-hungarian empire in 1914</code>	<code>[0.0262451171875, 0.0556640625, -0.0, -0.03076171875, -0.05712890625, ...]</code>
  <code>Instruct: Given a web search query, retrieve relevant passages that answer the query<br>Query:who played tommy in coward of the county</code>	<code>[-0.00848388671875, -0.02294921875, -0.00182342529296875, 0.060546875, -0.021240234375, ...]</code>
  <code>Vertebra The vertebral arch is formed by pedicles and laminae. Two pedicles extend from the sides of the vertebral body to join the body to the arch. The pedicles are short thick processes that extend, one from each side, posteriorly, from the junctions of the posteriolateral surfaces of the centrum, on its upper surface. From each pedicle a broad plate, a lamina, projects backwards and medialwards to join and complete the vertebral arch and form the posterior border of the vertebral foramen, which completes the triangle of the vertebral foramen.[6] The upper surfaces of the laminae are rough to give attachment to the ligamenta flava. These ligaments connect the laminae of adjacent vertebra along the length of the spine from the level of the second cervical vertebra. Above and below the pedicles are shallow depressions called vertebral notches (superior and inferior). When the vertebrae articulate the notches align with those on adjacent vertebrae and these form the openings of the int...</code>	<code>[0.062255859375, -0.005706787109375, -0.009765625, 0.035400390625, -0.0125732421875, ...]</code>

• Loss: <code>MSELoss</code>

gooaq

• Dataset: gooaq at b089f72
• Size: 3,000 evaluation samples
• Columns: <code>text</code> and <code>label</code>
• Approximate statistics based on the first 1000 samples:

  	text	label
  type	string	list
  details	<ul><li>min: 10 tokens</li><li>mean: 43.88 tokens</li><li>max: 117 tokens</li></ul>	<ul><li>size: 1024 elements</li></ul>

• Samples:

  text	label
  <code>Instruct: Given a web search query, retrieve relevant passages that answer the query<br>Query:what essential oils are soothing?</code>	<code>[-0.025146484375, 0.06591796875, -0.0025634765625, 0.0732421875, -0.046630859375, ...]</code>
  <code>Titles of books should be underlined or put in italics . (Titles of stories, essays and poems are in "quotation marks.") Refer to the text specifically as a novel, story, essay, memoir, or poem, depending on what it is.</code>	<code>[-0.006988525390625, -0.050537109375, -0.007476806640625, -0.07177734375, -0.049560546875, ...]</code>
  <code>Dakine Cyclone Wet/Dry 32L Backpack. Born from the legacy of our most iconic surf pack, the Cyclone Collection is a family of super-technical and durable wet/dry packs and bags.</code>	<code>[0.0016632080078125, 0.04150390625, -0.01324462890625, 0.0234375, 0.03173828125, ...]</code>

• Loss: <code>MSELoss</code>

wikipedia

• Dataset: wikipedia at 4a0972d
• Size: 3,000 evaluation samples
• Columns: <code>text</code> and <code>label</code>
• Approximate statistics based on the first 1000 samples:

  	text	label
  type	string	list
  details	<ul><li>min: 5 tokens</li><li>mean: 28.1 tokens</li><li>max: 105 tokens</li></ul>	<ul><li>size: 1024 elements</li></ul>

• Samples:

  text	label
  <code>The daughter of Vice-admiral George Davies and Julia Hume, she spent her younger years on board the ship he was stationed, the Griper.</code>	<code>[0.0361328125, 0.01904296875, -0.003662109375, 0.0247802734375, 0.0140380859375, ...]</code>
  <code>The impetus for the project began when Amalgamated Dynamics, hired to provide the practical effects for The Thing, a prequel to John Carpenter's 1982 classic film-renowned for its almost exclusive use of practical effects-became disillusioned upon discovering the theatrical release had the bulk of their effects digitally replaced with computer-generated imagery.</code>	<code>[-0.0106201171875, -0.0439453125, -0.01104736328125, 0.00946044921875, 0.0322265625, ...]</code>
  <code>Lost Angeles, his second feature film, starring Joelle Carter and Kelly Blatz, had its world premiere at the Oldenburg International Film Festival in 2012.</code>	<code>[0.0272216796875, 0.0263671875, -0.007110595703125, 0.0294189453125, 0.01129150390625, ...]</code>

• Loss: <code>MSELoss</code>

Training Hyperparameters

Non-Default Hyperparameters

• eval_strategy: steps
• per_device_train_batch_size: 16
• per_device_eval_batch_size: 16
• learning_rate: 0.0001
• num_train_epochs: 1
• warmup_ratio: 0.1
• bf16: True
• load_best_model_at_end: True

All Hyperparameters

<details><summary>Click to expand</summary>

• overwrite_output_dir: False
• do_predict: False
• eval_strategy: steps
• prediction_loss_only: True
• per_device_train_batch_size: 16
• per_device_eval_batch_size: 16
• 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: 0.0001
• 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: linear
• 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: 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}
• 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: proportional

</details>

Training Logs

• The bold row denotes the saved checkpoint.

Framework Versions

• Python: 3.11.10
• Sentence Transformers: 4.2.0.dev0
• Transformers: 4.51.2
• PyTorch: 2.5.1+cu124
• Accelerate: 1.5.2
• Datasets: 3.5.0
• Tokenizers: 0.21.0

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",
}

MSELoss

@inproceedings{reimers-2020-multilingual-sentence-bert,
    title = "Making Monolingual Sentence Embeddings Multilingual using Knowledge Distillation",
    author = "Reimers, Nils and Gurevych, Iryna",
    booktitle = "Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing",
    month = "11",
    year = "2020",
    publisher = "Association for Computational Linguistics",
    url = "https://arxiv.org/abs/2004.09813",
}

<!-- ## Glossary *Clearly define terms in order to be accessible across audiences.* --> <!-- ## Model Card Authors *Lists the people who create the model card, providing recognition and accountability for the detailed work that goes into its construction.* --> <!-- ## Model Card Contact *Provides a way for people who have updates to the Model Card, suggestions, or questions, to contact the Model Card authors.* -->