KaiserML/RhetoriBERT

sentence-transformers/static-retrieval-mrl-en-v1

This is a sentence-transformers model finetuned from nomic-ai/nomic-embed-text-v1.5 on the sci_gen_colbert_triplets dataset. It maps sentences from academic texts to a 768-dimensional dense vector space based on their rhetorical function (summarizing results, expressing limitations etc.) and can be used for functional textual similarity, limitations analysis, rhetorical function classification, clustering and more.

Model Details

Model Description

• Model Type: Sentence Transformer
• Base model: nomic-ai/nomic-embed-text-v1.5
• Maximum Sequence Length: 8192 tokens
• Output Dimensionality: 768 dimensions
• Similarity Function: Cosine Similarity
• Training Dataset:
  • sci_gen_colbert_triplets
• 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': 8192, 'do_lower_case': False}) with Transformer model: NomicBertModel 
  (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("Corran/SciGenNomicEmbed")
# Run inference
sentences = [
    'Surveys and interviews: Introducing excerpts from interview data',
    "Through surveys and interviews, multiliterate teachers expressed a shared belief in the importance of fostering students' ability to navigate multiple discourse communities.",
    'The authors employ a constructivist approach to learning, where students build knowledge through active engagement with multimedia texts and collaborative discussions.',
]
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

Information Retrieval

• Dataset: SciGen-Eval-Set
• Evaluated with InformationRetrievalEvaluator

Metric	Value
cosine_accuracy@1	0.9
cosine_accuracy@3	0.9452
cosine_accuracy@5	0.9642
cosine_accuracy@10	0.9853
cosine_precision@1	0.9
cosine_precision@3	0.3151
cosine_precision@5	0.1928
cosine_precision@10	0.0985
cosine_recall@1	0.9
cosine_recall@3	0.9452
cosine_recall@5	0.9642
cosine_recall@10	0.9853
cosine_ndcg@10	0.9415
cosine_mrr@10	0.9276
cosine_map@100	0.9284

Training Details

Training Dataset

sci_gen_colbert_triplets

• Dataset: sci_gen_colbert_triplets at 44071bd
• Size: 35,934 training samples
• Columns: query, positive, and negative
• Approximate statistics based on the first 1000 samples:

  	query	positive	negative
  type	string	string	string
  details	min: 5 tokens mean: 10.24 tokens max: 23 tokens	min: 2 tokens mean: 39.86 tokens max: 80 tokens	min: 18 tokens mean: 40.41 tokens max: 88 tokens

• Samples:

  query	positive	negative
  Previous research: highlighting negative outcomes	Despite the widespread use of seniority-based wage systems in labor contracts, previous research has highlighted their negative outcomes, such as inefficiencies and demotivating effects on workers.	This paper, published in 1974, was among the first to establish the importance of rank-order tournaments as optimal labor contracts in microeconomics.
  Synthesising sources: contrasting evidence or ideas	Despite the observed chronic enterocolitis in Interleukin-10-deficient mice, some studies suggest that this cytokine plays a protective role in intestinal inflammation in humans (Kurimoto et al., 2001).	Chronic enterocolitis developed in Interleukin-10-deficient mice, characterized by inflammatory cell infiltration, epithelial damage, and increased production of pro-inflammatory cytokines.
  Previous research: Approaches taken	Previous research on measuring patient-relevant outcomes in osteoarthritis has primarily relied on self-reported measures, such as the Western Ontario and McMaster Universities Arthritis Index (WOMAC) (Bellamy et al., 1988).	The WOMAC (Western Ontario and McMaster Universities Osteoarthritis Index) questionnaire has been widely used in physical therapy research to assess the impact of antirheumatic drug therapy on patient-reported outcomes in individuals with hip or knee osteoarthritis.

• Loss: MatryoshkaLoss with these parameters:

  {
      "loss": "MultipleNegativesRankingLoss",
      "matryoshka_dims": [
          768,
          384,
          256,
          128,
          64
      ],
      "matryoshka_weights": [
          1,
          1,
          1,
          1,
          1
      ],
      "n_dims_per_step": -1
  }
  

Evaluation Dataset

sci_gen_colbert_triplets

• Dataset: sci_gen_colbert_triplets at 44071bd
• Size: 4,492 evaluation samples
• Columns: query, positive, and negative
• Approximate statistics based on the first 1000 samples:

  	query	positive	negative
  type	string	string	string
  details	min: 5 tokens mean: 10.23 tokens max: 23 tokens	min: 18 tokens mean: 39.83 tokens max: 84 tokens	min: 8 tokens mean: 39.89 tokens max: 84 tokens

• Samples:

  query	positive	negative
  Providing background information: reference to the purpose of the study	This study aimed to investigate the impact of socioeconomic status on child development, specifically focusing on cognitive, language, and social-emotional domains.	Children from high socioeconomic status families showed significantly higher IQ scores (M = 112.5, SD = 5.6) compared to children from low socioeconomic status families (M = 104.3, SD = 6.2) in the verbal IQ subtest.
  Providing background information: reference to the literature	According to previous studies using WinGX suite for small-molecule single-crystal crystallography, the optimization of crystal structures leads to improved accuracy in determining atomic coordinates.	This paper describes the WinGX suite, a powerful tool for small-molecule single-crystal crystallography that significantly advances the field of crystallography by streamlining data collection and analysis.
  General comments on the relevant literature	Polymer brushes have gained significant attention in the field of polymer science due to their unique properties, such as controlled thickness, high surface density, and tunable interfacial properties.	Despite previous reports suggesting that polymer brushes with short grafting densities exhibit poorer performance in terms of adhesion and stability compared to those with higher grafting densities (Liu et al., 2010), our results indicate that the opposite is true for certain types of polymer brushes.

• Loss: MatryoshkaLoss with these parameters:

  {
      "loss": "MultipleNegativesRankingLoss",
      "matryoshka_dims": [
          768,
          384,
          256,
          128,
          64
      ],
      "matryoshka_weights": [
          1,
          1,
          1,
          1,
          1
      ],
      "n_dims_per_step": -1
  }
  

Training Hyperparameters

Non-Default Hyperparameters

• eval_strategy: steps
• per_device_train_batch_size: 256
• per_device_eval_batch_size: 256
• learning_rate: 2e-05
• num_train_epochs: 10
• warmup_ratio: 0.1
• fp16: True
• load_best_model_at_end: 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: 256
• per_device_eval_batch_size: 256
• 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: 2e-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: 10
• 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: 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: 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}
• 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: batch_sampler
• multi_dataset_batch_sampler: proportional

Training Logs

Epoch	Step	Training Loss	Validation Loss	SciGen-Eval-Set_cosine_ndcg@10
0	0	-	-	0.1744
0.1418	20	31.1056	29.9614	0.2010
0.2837	40	28.3636	25.9021	0.3552
0.4255	60	23.8421	21.4941	0.4817
0.5674	80	20.2484	19.1669	0.5793
0.7092	100	18.6804	18.0565	0.6219
0.8511	120	17.7705	17.3231	0.6564
0.9929	140	17.1951	16.8645	0.6723
1.1348	160	16.1046	16.3714	0.6918
1.2766	180	16.0491	16.0427	0.7063
1.4184	200	15.4859	15.6624	0.7240
1.5603	220	15.3239	15.4609	0.7341
1.7021	240	14.9202	15.1556	0.7414
1.8440	260	14.7176	14.8438	0.7584
1.9858	280	14.5036	14.5248	0.7718
2.1277	300	12.8219	14.4285	0.7860
2.2695	320	12.9107	14.1397	0.7927
2.4113	340	12.6728	13.8471	0.8092
2.5532	360	12.4097	13.6623	0.8160
2.6950	380	12.3039	13.4078	0.8264
2.8369	400	12.121	13.1426	0.8382
2.9787	420	12.0307	12.7989	0.8520
3.1206	440	10.4306	12.7893	0.8566
3.2624	460	10.5238	12.7036	0.8681
3.4043	480	10.3648	12.5674	0.8783
3.5461	500	10.4774	12.3069	0.8794
3.6879	520	10.4965	12.0965	0.8837
3.8298	540	10.4085	12.0368	0.8868
3.9716	560	10.2881	11.9063	0.8946
4.1135	580	9.1967	11.9930	0.8970
4.2553	600	9.3798	11.8936	0.9047
4.3972	620	9.3375	11.7678	0.9118
4.5390	640	9.2483	11.7572	0.9078
4.6809	660	9.3736	11.6011	0.9174
4.8227	680	9.3427	11.5383	0.9197
4.9645	700	9.3935	11.4293	0.9242
5.1064	720	8.5631	11.5119	0.9294
5.2482	740	8.6057	11.5173	0.9255
5.3901	760	8.6059	11.5421	0.9263
5.5319	780	8.8488	11.3879	0.9304
5.6738	800	8.7855	11.3523	0.9320
5.8156	820	8.7525	11.2572	0.9331
5.9574	840	8.8674	11.1829	0.9329
6.0993	860	8.0564	11.3401	0.9367
6.2411	880	8.1608	11.3323	0.9370
6.3830	900	8.2702	11.3146	0.9370
6.5248	920	8.3711	11.2561	0.9372
6.6667	940	8.421	11.2558	0.9354
6.8085	960	8.4125	11.1738	0.9384
6.9504	980	8.42	11.0996	0.9415

Framework Versions

• Python: 3.11.11
• Sentence Transformers: 3.3.1
• Transformers: 4.47.1
• PyTorch: 2.5.1+cu121
• Accelerate: 1.2.1
• 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",
}

MatryoshkaLoss

@misc{kusupati2024matryoshka,
    title={Matryoshka Representation Learning},
    author={Aditya Kusupati and Gantavya Bhatt and Aniket Rege and Matthew Wallingford and Aditya Sinha and Vivek Ramanujan and William Howard-Snyder and Kaifeng Chen and Sham Kakade and Prateek Jain and Ali Farhadi},
    year={2024},
    eprint={2205.13147},
    archivePrefix={arXiv},
    primaryClass={cs.LG}
}

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}
}