bobox/DeBERTaV3-small-ST-AdaptiveLayer-Norm-ep2

SentenceTransformer based on microsoft/deberta-v3-small

This is a sentence-transformers model finetuned from microsoft/deberta-v3-small on the stanfordnlp/snli dataset. It maps sentences & paragraphs to a 768-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: microsoft/deberta-v3-small
• Maximum Sequence Length: 512 tokens
• Output Dimensionality: 768 tokens
• Similarity Function: Cosine Similarity
• Training Dataset:
  • stanfordnlp/snli
• Language: en

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: DebertaV2Model 
  (1): Pooling({'word_embedding_dimension': 768, '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})
)

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("bobox/DeBERTaV3-small-ST-AdaptiveLayer-Norm-ep2")
# Run inference
sentences = [
    'First Lady Laura Bush at podium, in front of seated audience, at the White House Conference on Global Literacy.',
    'The former First Lady is at the podium for a conference.',
    'This person is going to the waterfall',
]
embeddings = model.encode(sentences)
print(embeddings.shape)
# [3, 768]

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

Evaluation

Metrics

Binary Classification

• Evaluated with BinaryClassificationEvaluator

Metric	Value
cosine_accuracy	0.6651
cosine_accuracy_threshold	0.6879
cosine_f1	0.7077
cosine_f1_threshold	0.6305
cosine_precision	0.6223
cosine_recall	0.8204
cosine_ap	0.7058
dot_accuracy	0.6313
dot_accuracy_threshold	135.985
dot_f1	0.6997
dot_f1_threshold	115.5461
dot_precision	0.58
dot_recall	0.8817
dot_ap	0.6555
manhattan_accuracy	0.6708
manhattan_accuracy_threshold	219.3239
manhattan_f1	0.712
manhattan_f1_threshold	262.3147
manhattan_precision	0.6062
manhattan_recall	0.8624
manhattan_ap	0.7135
euclidean_accuracy	0.6653
euclidean_accuracy_threshold	11.5068
euclidean_f1	0.708
euclidean_f1_threshold	12.4785
euclidean_precision	0.6209
euclidean_recall	0.8236
euclidean_ap	0.709
max_accuracy	0.6708
max_accuracy_threshold	219.3239
max_f1	0.712
max_f1_threshold	262.3147
max_precision	0.6223
max_recall	0.8817
max_ap	0.7135

Training Details

Training Dataset

stanfordnlp/snli

• Dataset: stanfordnlp/snli at cdb5c3d
• Size: 67,190 training samples
• Columns: sentence1, sentence2, and label
• Approximate statistics based on the first 1000 samples:

  	sentence1	sentence2	label
  type	string	string	int
  details	min: 4 tokens mean: 21.19 tokens max: 133 tokens	min: 4 tokens mean: 11.77 tokens max: 49 tokens	0: 100.00%

• Samples:

  sentence1	sentence2	label
  Without a placebo group, we still won't know if any of the treatments are better than nothing and therefore worth giving.	It is necessary to use a controlled method to ensure the treatments are worthwhile.	0
  It was conducted in silence.	It was done silently.	0
  oh Lewisville any decent food in your cafeteria up there	Is there any decent food in your cafeteria up there in Lewisville?	0

• Loss: AdaptiveLayerLoss with these parameters:

  {
      "loss": "MultipleNegativesRankingLoss",
      "n_layers_per_step": 1,
      "last_layer_weight": 1,
      "prior_layers_weight": 0.05,
      "kl_div_weight": 2,
      "kl_temperature": 0.9
  }
  

Evaluation Dataset

stanfordnlp/snli

• Dataset: stanfordnlp/snli at cdb5c3d
• Size: 6,626 evaluation samples
• Columns: premise, hypothesis, and label
• Approximate statistics based on the first 1000 samples:

  	premise	hypothesis	label
  type	string	string	int
  details	min: 6 tokens mean: 17.28 tokens max: 59 tokens	min: 4 tokens mean: 10.53 tokens max: 32 tokens	0: ~48.70% 1: ~51.30%

• Samples:

  premise	hypothesis	label
  This church choir sings to the masses as they sing joyous songs from the book at a church.	The church has cracks in the ceiling.	0
  This church choir sings to the masses as they sing joyous songs from the book at a church.	The church is filled with song.	1
  A woman with a green headscarf, blue shirt and a very big grin.	The woman is young.	0

• Loss: AdaptiveLayerLoss with these parameters:

  {
      "loss": "MultipleNegativesRankingLoss",
      "n_layers_per_step": 1,
      "last_layer_weight": 1,
      "prior_layers_weight": 0.05,
      "kl_div_weight": 2,
      "kl_temperature": 0.9
  }
  

Training Hyperparameters

Non-Default Hyperparameters

• eval_strategy: steps
• per_device_train_batch_size: 45
• per_device_eval_batch_size: 22
• learning_rate: 3e-06
• weight_decay: 1e-09
• num_train_epochs: 2
• lr_scheduler_type: cosine
• warmup_ratio: 0.5
• save_safetensors: False
• fp16: True
• push_to_hub: True
• hub_model_id: bobox/DeBERTaV3-small-ST-AdaptiveLayer-Norm-ep2-checkpoints
• hub_strategy: checkpoint
• 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: 45
• per_device_eval_batch_size: 22
• per_gpu_train_batch_size: None
• per_gpu_eval_batch_size: None
• gradient_accumulation_steps: 1
• eval_accumulation_steps: None
• learning_rate: 3e-06
• weight_decay: 1e-09
• adam_beta1: 0.9
• adam_beta2: 0.999
• adam_epsilon: 1e-08
• max_grad_norm: 1.0
• num_train_epochs: 2
• max_steps: -1
• lr_scheduler_type: cosine
• lr_scheduler_kwargs: {}
• warmup_ratio: 0.5
• warmup_steps: 0
• log_level: passive
• log_level_replica: warning
• log_on_each_node: True
• logging_nan_inf_filter: True
• save_safetensors: False
• 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}
• 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: True
• resume_from_checkpoint: None
• hub_model_id: bobox/DeBERTaV3-small-ST-AdaptiveLayer-Norm-ep2-checkpoints
• hub_strategy: checkpoint
• 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
• batch_sampler: no_duplicates
• multi_dataset_batch_sampler: proportional

Training Logs

Epoch	Step	Training Loss	loss	max_ap
0.1004	150	4.5827	-	-
0.2001	299	-	3.5735	0.6133
0.2008	300	3.5451	-	-
0.3012	450	2.9066	-	-
0.4003	598	-	2.8785	0.6561
0.4016	600	2.5141	-	-
0.5020	750	2.0248	-	-
0.6004	897	-	2.1300	0.6917
0.6024	900	1.6782	-	-
0.7028	1050	1.4187	-	-
0.8005	1196	-	1.7111	0.7051
0.8032	1200	1.2446	-	-
0.9036	1350	1.1078	-	-
1.0007	1495	-	1.4859	0.7108
1.0040	1500	0.9827	-	-
1.1044	1650	0.9335	-	-
1.2008	1794	-	1.3516	0.7121
1.2048	1800	0.8595	-	-
1.3052	1950	0.8362	-	-
1.4009	2093	-	1.2659	0.7147
1.4056	2100	0.8167	-	-
1.5060	2250	0.7695	-	-
1.6011	2392	-	1.2218	0.7135
1.6064	2400	0.7544	-	-
1.7068	2550	0.7625	-	-
1.8012	2691	-	1.2073	0.7135
1.8072	2700	0.7366	-	-
1.9076	2850	0.7348	-	-

Framework Versions

• Python: 3.10.13
• Sentence Transformers: 3.0.1
• Transformers: 4.41.2
• PyTorch: 2.1.2
• Accelerate: 0.30.1
• Datasets: 2.19.2
• Tokenizers: 0.19.1

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",
}

AdaptiveLayerLoss

@misc{li20242d,
    title={2D Matryoshka Sentence Embeddings}, 
    author={Xianming Li and Zongxi Li and Jing Li and Haoran Xie and Qing Li},
    year={2024},
    eprint={2402.14776},
    archivePrefix={arXiv},
    primaryClass={cs.CL}
}

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}
}