redis/unified-negatives

SentenceTransformer based on thenlper/gte-small

This is a sentence-transformers model finetuned from thenlper/gte-small. It maps sentences & paragraphs to a 384-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: thenlper/gte-small
• Maximum Sequence Length: 128 tokens
• Output Dimensionality: 384 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': 128, 'do_lower_case': False, 'architecture': 'BertModel'})
  (1): Pooling({'word_embedding_dimension': 384, '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("redis/unified-negatives")
# Run inference
sentences = [
    'What is the fastest way to get a PAN card within India?',
    'What is the fastest way to get a PAN card within India?',
    'What is the fastest way to get a PAN card outside India?',
]
embeddings = model.encode(sentences)
print(embeddings.shape)
# [3, 384]

# Get the similarity scores for the embeddings
similarities = model.similarity(embeddings, embeddings)
print(similarities)
# tensor([[1.0000, 1.0000, 0.2943],
#         [1.0000, 1.0000, 0.2943],
#         [0.2943, 0.2943, 1.0000]])

Evaluation

Metrics

Information Retrieval

• Datasets: NanoMSMARCO and NanoNQ
• Evaluated with InformationRetrievalEvaluator

Metric	NanoMSMARCO	NanoNQ
cosine_accuracy@1	0.28	0.32
cosine_accuracy@3	0.48	0.6
cosine_accuracy@5	0.52	0.66
cosine_accuracy@10	0.58	0.74
cosine_precision@1	0.28	0.32
cosine_precision@3	0.16	0.2
cosine_precision@5	0.104	0.132
cosine_precision@10	0.058	0.074
cosine_recall@1	0.28	0.3
cosine_recall@3	0.48	0.55
cosine_recall@5	0.52	0.61
cosine_recall@10	0.58	0.68
cosine_ndcg@10	0.4281	0.5109
cosine_mrr@10	0.3795	0.4792
cosine_map@100	0.3902	0.4526

Nano BEIR

• Dataset: NanoBEIR_mean
• Evaluated with NanoBEIREvaluator with these parameters:

  {
      "dataset_names": [
          "msmarco",
          "nq"
      ],
      "dataset_id": "lightonai/NanoBEIR-en"
  }
  

Metric	Value
cosine_accuracy@1	0.3
cosine_accuracy@3	0.54
cosine_accuracy@5	0.59
cosine_accuracy@10	0.66
cosine_precision@1	0.3
cosine_precision@3	0.18
cosine_precision@5	0.118
cosine_precision@10	0.066
cosine_recall@1	0.29
cosine_recall@3	0.515
cosine_recall@5	0.565
cosine_recall@10	0.63
cosine_ndcg@10	0.4695
cosine_mrr@10	0.4294
cosine_map@100	0.4214

Training Details

Training Dataset

Unnamed Dataset

• Size: 21,470 training samples
• Columns: anchor, positive, and negative
• Approximate statistics based on the first 1000 samples:

  	anchor	positive	negative
  type	string	string	string
  details	min: 5 tokens mean: 19.91 tokens max: 101 tokens	min: 5 tokens mean: 19.91 tokens max: 101 tokens	min: 4 tokens mean: 19.91 tokens max: 101 tokens

• Samples:

  anchor	positive	negative
  The pale coloration provides camouflage for the beetle on the light sand.	The pale coloration provides camouflage for the beetle on the light sand.	The pale coloration helps the beetle stand out on the light sand.
  It is found from Fennoscandinavia to the Pyrenees , Italy and Greece and from Britain to Russia and Ukraine .	It is found from Fennoscandinavia to the Pyrenees , Italy and Greece and from Britain to Russia and Ukraine .	It is located from Fennoscandinavia to the Pyrenees , Great Britain and Greece and from Italy to Russia and Ukraine .
  Is Swami Vivekananda's speech at parliament of world's religions, Chicago overrated in Chicago?	Is Swami Vivekananda's speech at parliament of world's religions, Chicago overrated in Chicago?	Is Swami Vivekananda's speech at parliament of world's religions, Chicago overrated outside Chicago?

• Loss: MultipleNegativesRankingLoss with these parameters:

  {
      "scale": 7.0,
      "similarity_fct": "cos_sim",
      "gather_across_devices": false
  }
  

Evaluation Dataset

Unnamed Dataset

• Size: 2,386 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: 19.42 tokens max: 74 tokens	min: 6 tokens mean: 19.42 tokens max: 74 tokens	min: 6 tokens mean: 19.41 tokens max: 74 tokens

• Samples:

  anchor	positive	negative
  He died at Fort Edward on August 18 , 1861 , and was buried at the Union Cemetery in Sandy Hill .	He died at Fort Edward on August 18 , 1861 , and was buried at the Union Cemetery in Sandy Hill .	He died at Sandy Hill on August 18 , 1861 , and was buried at the Union Cemetery in Fort Edward .
  It was this cooperation which led to the development of the satellite AIS system.	It was this cooperation which led to the development of the satellite AIS system.	It was this cooperation which led to the halting of development of the satellite AIS system.
  What is the best field of engineering on campus?	What is the best field of engineering on campus?	What is the best field of engineering off campus?

• Loss: MultipleNegativesRankingLoss with these parameters:

  {
      "scale": 7.0,
      "similarity_fct": "cos_sim",
      "gather_across_devices": false
  }
  

Training Hyperparameters

Non-Default Hyperparameters

• eval_strategy: steps
• per_device_train_batch_size: 128
• per_device_eval_batch_size: 128
• learning_rate: 1e-06
• weight_decay: 0.001
• max_steps: 3000
• warmup_ratio: 0.1
• fp16: True
• dataloader_drop_last: True
• dataloader_num_workers: 1
• dataloader_prefetch_factor: 1
• load_best_model_at_end: True
• optim: adamw_torch
• ddp_find_unused_parameters: False
• push_to_hub: True
• hub_model_id: redis/unified-negatives
• eval_on_start: True

All Hyperparameters

Click to expand

• overwrite_output_dir: False
• do_predict: False
• eval_strategy: steps
• prediction_loss_only: True
• per_device_train_batch_size: 128
• per_device_eval_batch_size: 128
• 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: 1e-06
• weight_decay: 0.001
• adam_beta1: 0.9
• adam_beta2: 0.999
• adam_epsilon: 1e-08
• max_grad_norm: 1.0
• num_train_epochs: 3.0
• max_steps: 3000
• 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
• 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: True
• dataloader_num_workers: 1
• dataloader_prefetch_factor: 1
• 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: False
• 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: True
• resume_from_checkpoint: None
• hub_model_id: redis/unified-negatives
• 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: True
• use_liger_kernel: False
• liger_kernel_config: None
• eval_use_gather_object: False
• average_tokens_across_devices: True
• prompts: None
• batch_sampler: batch_sampler
• multi_dataset_batch_sampler: proportional
• router_mapping: {}
• learning_rate_mapping: {}

Training Logs

Epoch	Step	Training Loss	Validation Loss	NanoMSMARCO_cosine_ndcg@10	NanoNQ_cosine_ndcg@10	NanoBEIR_mean_cosine_ndcg@10
0	0	-	3.6734	0.6259	0.6583	0.6421
1.4970	250	3.8677	3.3900	0.6334	0.6510	0.6422
2.9940	500	3.188	1.8654	0.5772	0.6252	0.6012
4.4910	750	1.4714	0.6890	0.4032	0.5437	0.4735
5.9880	1000	0.8535	0.5511	0.3617	0.5197	0.4407
7.4850	1250	0.7547	0.5268	0.3469	0.5346	0.4407
8.9820	1500	0.716	0.5123	0.3684	0.5223	0.4454
10.4790	1750	0.6939	0.5039	0.3846	0.5179	0.4512
11.9760	2000	0.6789	0.4986	0.4120	0.5280	0.4700
13.4731	2250	0.6681	0.4953	0.4148	0.5189	0.4669
14.9701	2500	0.662	0.4918	0.4224	0.5109	0.4666
16.4671	2750	0.6575	0.4905	0.4224	0.5109	0.4666
17.9641	3000	0.6555	0.4900	0.4281	0.5109	0.4695

Framework Versions

• Python: 3.10.18
• Sentence Transformers: 5.2.0
• Transformers: 4.57.3
• PyTorch: 2.9.1+cu128
• Accelerate: 1.12.0
• Datasets: 2.21.0
• Tokenizers: 0.22.1

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}
}