ancc/Velvet-2B-embedding-news

SentenceTransformer based on Almawave/Velvet-2B

This is a sentence-transformers model finetuned from Almawave/Velvet-2B. It maps sentences & paragraphs to a 2048-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: Almawave/Velvet-2B
• Maximum Sequence Length: 32768 tokens
• Output Dimensionality: 2048 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': 32768, 'do_lower_case': False}) with Transformer model: MistralModel 
  (1): Pooling({'word_embedding_dimension': 2048, '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("ancc/Velvet-2B-embedding-news")
# Run inference
sentences = [
    "NBA, dopo l’addio a Schroeder i Nets promuovono Simmons in quintetto: l'idea è correre",
    'Italia Viva e +Europa non parteciperanno alle elezioni regionali in Liguria',
    'Achille Costacurta rivela: «Sono stato rinchiuso per un anno e sette mesi in un centro penale',
]
embeddings = model.encode(sentences)
print(embeddings.shape)
# [3, 2048]

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

Evaluation

Metrics

Paraphrase Mining

• Evaluated with ParaphraseMiningEvaluator

Metric	Value
average_precision	0.5284
f1	0.5502
precision	0.5568
recall	0.5439
threshold	0.931

Training Details

Training Dataset

Unnamed Dataset

• Size: 152,913 training samples
• Columns: sentence and label
• Approximate statistics based on the first 1000 samples:

  	sentence	label
  type	string	int
  details	min: 3 tokens mean: 19.14 tokens max: 47 tokens	0: ~0.50% 1: ~0.50% 2: ~0.50% 3: ~0.50% 4: ~0.50% 5: ~0.20% 6: ~0.50% 7: ~0.50% 8: ~0.20% 9: ~0.50% 11: ~0.50% 12: ~0.50% 13: ~0.50% 15: ~0.50% 16: ~0.20% 17: ~0.30% 18: ~0.50% 19: ~0.40% 20: ~0.50% 21: ~0.50% 22: ~0.30% 23: ~0.50% 24: ~0.50% 25: ~0.50% 26: ~0.50% 27: ~0.50% 28: ~0.40% 29: ~0.50% 30: ~0.20% 31: ~0.30% 32: ~0.40% 33: ~0.50% 34: ~0.50% 36: ~0.50% 38: ~0.50% 39: ~0.50% 40: ~0.30% 41: ~0.20% 43: ~0.50% 44: ~0.50% 45: ~0.50% 46: ~0.40% 47: ~0.30% 48: ~0.30% 49: ~0.50% 50: ~0.30% 51: ~0.50% 52: ~0.50% 53: ~0.50% 54: ~0.50% 55: ~0.20% 56: ~0.50% 57: ~0.50% 59: ~0.50% 60: ~0.40% 62: ~0.50% 63: ~0.20% 64: ~0.50% 65: ~0.40% 66: ~0.50% 67: ~0.40% 68: ~0.50% 69: ~0.50% 70: ~0.50% 71: ~0.20% 72: ~0.50% 73: ~0.20% 74: ~0.50% 75: ~0.50% 77: ~0.50% 78: ~0.30% 79: ~0.50% 80: ~0.40% 81: ~0.30% 82: ~0.20% 83: ~0.40% 84: ~0.30% 85: ~0.50% 87: ~0.50% 88: ~0.20% 90: ~0.50% 91: ~0.20% 92: ~0.50% 93: ~0.40% 94: ~0.50% 95: ~0.50% 97: ~0.50% 98: ~0.40% 99: ~0.50% 100: ~0.30% 101: ~0.50% 103: ~0.20% 104: ~0.50% 106: ~0.40% 107: ~0.20% 108: ~0.40% 109: ~0.30% 110: ~0.50% 111: ~0.40% 112: ~0.50% 113: ~0.30% 115: ~0.30% 116: ~0.50% 117: ~0.50% 118: ~0.30% 120: ~0.30% 121: ~0.50% 122: ~0.20% 123: ~0.40% 124: ~0.30% 125: ~0.40% 126: ~0.50% 127: ~0.50% 128: ~0.50% 129: ~0.50% 130: ~0.50% 131: ~0.50% 132: ~0.40% 133: ~0.50% 134: ~0.30% 135: ~0.50% 136: ~0.30% 137: ~0.30% 138: ~0.50% 139: ~0.50% 140: ~0.20% 141: ~0.50% 142: ~0.50% 143: ~0.50% 144: ~0.30% 145: ~0.30% 146: ~0.30% 147: ~0.50% 148: ~0.30% 149: ~0.20% 150: ~0.50% 151: ~0.30% 152: ~0.20% 153: ~0.50% 154: ~0.50% 155: ~0.40% 156: ~0.20% 158: ~0.50% 159: ~0.50% 160: ~0.50% 161: ~0.40% 162: ~0.20% 163: ~0.40% 164: ~0.20% 165: ~0.50% 166: ~0.50% 167: ~0.50% 168: ~0.20% 169: ~0.50% 170: ~0.50% 171: ~0.20% 172: ~0.50% 173: ~0.50% 174: ~0.40% 175: ~0.50% 176: ~0.30% 178: ~0.50% 180: ~0.20% 182: ~0.50% 183: ~0.50% 184: ~0.30% 185: ~0.50% 187: ~0.20% 188: ~0.50% 189: ~0.50% 191: ~0.50% 192: ~0.50% 193: ~0.50% 194: ~0.20% 195: ~0.30% 196: ~0.50% 197: ~0.50% 199: ~0.50% 200: ~0.50% 201: ~0.20% 202: ~0.50% 203: ~0.50% 204: ~0.30% 205: ~0.50% 206: ~0.20% 207: ~0.40% 208: ~0.50% 209: ~0.30% 210: ~0.30% 211: ~0.50% 212: ~0.20% 213: ~0.50% 214: ~0.50% 215: ~0.40% 216: ~0.50% 217: ~0.40% 218: ~0.50% 219: ~0.50% 220: ~0.20% 221: ~0.50% 222: ~0.40% 223: ~0.50% 224: ~0.30% 225: ~0.40% 227: ~0.20% 228: ~0.30% 230: ~0.40% 231: ~0.40% 232: ~0.50% 233: ~0.50% 234: ~0.50% 235: ~0.50% 236: ~0.40% 237: ~0.50% 238: ~0.40% 239: ~0.50% 240: ~0.50% 241: ~0.50% 242: ~0.30% 243: ~0.50% 244: ~0.50% 245: ~0.50% 246: ~0.50% 247: ~0.20% 248: ~0.50% 249: ~0.50% 250: ~0.50% 251: ~0.50% 252: ~0.20% 253: ~0.50% 254: ~0.50% 255: ~0.30% 256: ~0.50% 257: ~0.50% 258: ~0.20% 259: ~0.50% 260: ~0.30% 261: ~0.50% 262: ~0.50% 263: ~0.20% 264: ~0.50% 265: ~0.20%

• Samples:

  sentence	label
  MERCATO LBA - Treviso, Giofrè: "Mercato in continua osservazione, vedremo..."	0
  Ky Bowman: Non sono soddisfatto delle mie performance	0
  LBA - Treviso, Giofrè: "Sabato la Reggiana, dobbiamo vincere. Punto"	0

• Loss: BatchAllTripletLoss

Evaluation Dataset

Unnamed Dataset

• Size: 9,310 evaluation samples
• Columns: sentence and label
• Approximate statistics based on the first 1000 samples:

  	sentence	label
  type	string	int
  details	min: 4 tokens mean: 19.04 tokens max: 78 tokens	0: ~0.50% 1: ~0.30% 2: ~0.50% 4: ~0.50% 5: ~0.20% 6: ~0.30% 7: ~0.50% 8: ~0.50% 9: ~0.50% 10: ~0.50% 12: ~0.40% 13: ~0.30% 14: ~0.30% 15: ~0.40% 16: ~0.20% 17: ~0.50% 18: ~0.40% 19: ~0.40% 20: ~0.50% 21: ~0.50% 22: ~0.50% 23: ~0.50% 24: ~0.20% 25: ~0.20% 26: ~0.50% 27: ~0.50% 28: ~0.50% 30: ~0.50% 31: ~0.20% 32: ~0.30% 33: ~0.50% 34: ~0.20% 35: ~0.50% 37: ~0.50% 38: ~0.20% 39: ~0.50% 40: ~0.50% 41: ~0.20% 42: ~0.30% 43: ~0.30% 44: ~0.50% 45: ~0.30% 46: ~0.50% 47: ~0.50% 48: ~0.30% 49: ~0.40% 50: ~0.30% 51: ~0.20% 52: ~0.50% 53: ~0.20% 54: ~0.30% 55: ~0.50% 56: ~0.50% 57: ~0.20% 58: ~0.50% 59: ~0.50% 60: ~0.50% 61: ~0.30% 62: ~0.30% 63: ~0.20% 65: ~0.50% 67: ~0.40% 68: ~0.50% 69: ~0.50% 70: ~0.20% 71: ~0.40% 72: ~0.50% 73: ~0.50% 74: ~0.30% 77: ~0.20% 78: ~0.50% 79: ~0.50% 80: ~0.20% 81: ~0.50% 82: ~0.20% 83: ~0.50% 84: ~0.50% 85: ~0.50% 89: ~0.20% 90: ~0.50% 91: ~0.50% 92: ~0.30% 93: ~0.40% 94: ~0.50% 95: ~0.30% 96: ~0.20% 97: ~0.50% 98: ~0.20% 99: ~0.50% 100: ~0.50% 101: ~0.50% 102: ~0.50% 103: ~0.30% 104: ~0.50% 105: ~0.50% 106: ~0.50% 107: ~0.50% 108: ~0.50% 109: ~0.50% 110: ~0.20% 111: ~0.50% 112: ~0.50% 113: ~0.20% 114: ~0.50% 115: ~0.40% 116: ~0.50% 117: ~0.20% 118: ~0.50% 120: ~0.50% 121: ~0.30% 122: ~0.50% 123: ~0.40% 124: ~0.20% 125: ~0.50% 126: ~0.20% 129: ~0.50% 130: ~0.30% 131: ~0.50% 132: ~0.40% 133: ~0.50% 134: ~0.50% 135: ~0.50% 136: ~0.20% 137: ~0.30% 138: ~0.50% 140: ~0.30% 141: ~0.30% 142: ~0.50% 143: ~0.40% 144: ~0.50% 145: ~0.50% 146: ~0.30% 147: ~0.20% 148: ~0.50% 149: ~0.40% 150: ~0.50% 151: ~0.20% 152: ~0.50% 154: ~0.20% 155: ~0.20% 156: ~0.50% 157: ~0.50% 158: ~0.50% 159: ~0.50% 160: ~0.20% 162: ~0.50% 163: ~0.20% 164: ~0.20% 165: ~0.20% 166: ~0.20% 167: ~0.30% 168: ~0.50% 169: ~0.50% 170: ~0.50% 171: ~0.40% 172: ~0.30% 173: ~0.50% 174: ~0.50% 175: ~0.30% 176: ~0.50% 178: ~0.50% 179: ~0.20% 180: ~0.50% 181: ~0.50% 182: ~0.30% 183: ~0.40% 184: ~0.20% 185: ~0.50% 186: ~0.50% 187: ~0.50% 188: ~0.20% 189: ~0.50% 191: ~0.20% 192: ~0.50% 193: ~0.20% 194: ~0.20% 195: ~0.40% 196: ~0.20% 197: ~0.20% 198: ~0.50% 199: ~0.40% 200: ~0.50% 201: ~0.50% 202: ~0.50% 203: ~0.40% 205: ~0.50% 206: ~0.50% 207: ~0.50% 208: ~0.50% 209: ~0.50% 210: ~0.50% 211: ~0.40% 212: ~0.50% 213: ~0.50% 214: ~0.20% 215: ~0.20% 216: ~0.50% 217: ~0.50% 218: ~0.50% 219: ~0.30% 220: ~0.50% 221: ~0.50% 223: ~0.50% 224: ~0.50% 225: ~0.50% 227: ~0.20% 228: ~0.50% 230: ~0.20% 231: ~0.50% 232: ~0.50% 234: ~0.30% 235: ~0.40% 236: ~0.50% 237: ~0.50% 238: ~0.20% 239: ~0.50% 240: ~0.20% 241: ~0.50% 242: ~0.20% 243: ~0.20% 244: ~0.20% 246: ~0.50% 247: ~0.20% 248: ~0.50% 249: ~0.50% 250: ~0.40% 251: ~0.50% 252: ~0.50% 253: ~0.20% 254: ~0.50% 255: ~0.50% 256: ~0.50% 257: ~0.50% 258: ~0.20% 259: ~0.20% 260: ~0.50% 261: ~0.50% 262: ~0.20% 263: ~0.40% 264: ~0.50% 265: ~0.50% 266: ~0.50% 267: ~0.50% 268: ~0.20% 269: ~0.20% 270: ~0.40% 272: ~0.30% 273: ~0.20% 274: ~0.50% 275: ~0.50% 276: ~0.40% 277: ~0.30% 278: ~0.30%

• Samples:

  sentence	label
  Supplenze: come funzionano i contratti fino al 31 dicembre 2024 e il calcolo del punteggio?	0
  Docente non abilitato assunto a tempo determinato da concorso PNRR1: in quale scuola "andrò a finire" se nella mia si perde un posto?	0
  Docenti non abilitati nominati dopo il 31 agosto da graduatorie pubblicate prima: otterranno sede di titolarità all’esito delle operazioni di mobilità [Chiarimenti]	0

• Loss: BatchAllTripletLoss

Training Hyperparameters

Non-Default Hyperparameters

• per_device_train_batch_size: 32
• num_train_epochs: 4
• lr_scheduler_type: cosine
• warmup_ratio: 0.25
• seed: 17
• data_seed: 17
• bf16: True
• batch_sampler: group_by_label

All Hyperparameters

Click to expand

• overwrite_output_dir: False
• do_predict: False
• eval_strategy: no
• prediction_loss_only: True
• per_device_train_batch_size: 32
• 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: 1.0
• num_train_epochs: 4
• max_steps: -1
• lr_scheduler_type: cosine
• lr_scheduler_kwargs: {}
• warmup_ratio: 0.25
• 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: 17
• data_seed: 17
• 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: 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: 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
• 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
• eval_on_start: False
• use_liger_kernel: False
• eval_use_gather_object: False
• average_tokens_across_devices: False
• prompts: None
• batch_sampler: group_by_label
• multi_dataset_batch_sampler: proportional

Training Logs

Epoch	Step	Training Loss	Validation Loss	average_precision
0.0002	1	0.0521	-	-
0.8002	3824	0.0195	-	-
1.7732	7648	0.0071	-	-
2.7462	11472	0.0052	-	-
3.7192	15296	0.0046	-	-
3.8272	15812	-	0.0011	0.5284

Framework Versions

• Python: 3.12.8
• Sentence Transformers: 3.4.1
• Transformers: 4.48.3
• PyTorch: 2.6.0+cu124
• Accelerate: 1.3.0
• Datasets: 3.2.0
• 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",
}

BatchAllTripletLoss

@misc{hermans2017defense,
    title={In Defense of the Triplet Loss for Person Re-Identification},
    author={Alexander Hermans and Lucas Beyer and Bastian Leibe},
    year={2017},
    eprint={1703.07737},
    archivePrefix={arXiv},
    primaryClass={cs.CV}
}