cristiano-sartori/bge-base-en-v1.5_finetuned

BGE base Financial Matryoshka

This is a sentence-transformers model finetuned from BAAI/bge-base-en-v1.5 on the json 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: BAAI/bge-base-en-v1.5
• Maximum Sequence Length: 512 tokens
• Output Dimensionality: 768 dimensions
• Similarity Function: Cosine Similarity
• Training Dataset:
  • json
• 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': 512, 'do_lower_case': True}) with Transformer model: BertModel 
  (1): Pooling({'word_embedding_dimension': 768, 'pooling_mode_cls_token': True, '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': 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("cristiano-sartori/bge-base-en-v1.5_finetuned")
# Run inference
sentences = [
    "I have the following code in my microcontroler program:\n\nint analogValue = ADCH;        // ADC Data Register\n\n//\n// Simple analog comparator. \n// If analogValue lower than threshold then toggle output high,\n// Otherwise toggle it low.\n//\nif ( analogValue > 128 ) {\n    PORTB = 0;                 // Port B Data Register\n} else {\n    PORTB = _BS( outputPin );  // Port B Data Register\n}\n\n\nWhere: \n\n\nADCH is the register that contains the value from the ADC\nPORTB is a dital output port that toggles an LED\n\n\nLooking at the resulting assembly code, I noticed that it is doing a 16 bit compare (lines 40-44) where strictly speaking only 8 bits would have been sufficient:\n\n40:   90 e0           ldi     r25, 0x00       ; 0\n42:   81 38           cpi     r24, 0x81       ; 129\n44:   91 05           cpc     r25, r1\n46:   14 f0           brlt    .+4             ; 0x4c <__SREG__+0xd>\n48:   18 ba           out     0x18, r1        ; PORTB\n4a:   f5 cf           rjmp    .-22            ; 0x36 <__CCP__+0x2>\n4c:   28 bb           out     0x18, r18       ; PORTB\n4e:   f3 cf           rjmp    .-26            ; 0x36 <__CCP__+0x2>\n\n\nI realize I declared analogValue as int, which indeed is 16 bit on AVR, but ...\n\nHow can I instruct the compiler to use 8 bit comparison? The Arduino IDE allows me to use byte, but avr-gcc by default doesn't.\n\nCheck this page for the complete program and its disassembled resulting code.\n\nEDIT1:\n\nChanging int to char changes the assembly code to:\n\n14:   11 24           eor     r1, r1          ; r1 = 0\n3e:   18 ba           out     0x18, r1        ; PORTB\n\n\nBasically skipping the test entirely.\n\nEDIT2: (Thnx: Wouter van Ooijen)\n\nChanging int to unsigned char changes the assembly code to:\n\n3c:   85 b1           in      r24, 0x05       ; ADCH\n3e:   ...\n40:   87 fd           sbrc    r24, 7          ; compare < 128 (well optimized)\n42:   02 c0           rjmp    .+4             ; 0x48 <__SREG__+0x9>\n44:   18 ba           out     0x18, r1        ; 24\n46:   f7 cf           rjmp    .-18            ; 0x36 <__CCP__+0x2>\n48:   98 bb           out     0x18, r25       ; 24\n4a:   f5 cf           rjmp    .-22            ; 0x36 <__CCP__+0x2>\n\n",
    'I actually think a better practice that avoids this architectural ambiguity is to include <stdint.h> then use declarative types like:\n\n\nuint8_t for unsigned 8-bit integers\nint8_t for signed 8-bit integers\nuint16_t for unsigned 16-bit integers\nuint32_t for unsigned 32-bit integers\n\n\nand so on...\n',
    'Entropy is a function of the distribution. That is, the process used to generate a byte stream is what has entropy, not the byte stream itself. If I give you the bits 1011, that could have anywhere from 0 to 4 bits of entropy; you have no way of knowing that value.\n\nHere is the definition of Shannon entropy. Let $X$ be a random variable that takes on the values $x_1,x_2,x_3,\\dots,x_n$. Then the Shannon entropy is defined as\n\n$$H(X) = -\\sum_{i=1}^{n} \\operatorname{Pr}[x_i] \\cdot \\log_2\\left(\\operatorname{Pr}[x_i]\\right)$$\n\nwhere $\\operatorname{Pr}[\\cdot]$ represents probability. Note that the definition is a function of a random variable (i.e., a distribution), not a particular value!\n\nSo what is the entropy in a single flip of a coin? Let $F$ be a random variable representing such. There are two events, heads and tails, each with probability $0.5$. So, the Shannon entropy of $F$ is:\n\n$$H(F) = -(0.5\\cdot\\log_2 0.5 + 0.5\\cdot\\log_2 0.5) = -(-0.5 + -0.5) = 1.$$\n\nThus, $F$ has exactly one bit of entropy, what we expected. \n\nSo, to find how much entropy is present in a byte stream, you need to know how the byte stream is generated and the entropy of any inputs (in the case of PRNGs). Recall that a deterministic algorithm cannot add entropy to an input, only take it away, so the entropy of all inputs to a deterministic algorithm is the maximum entropy possible in the output. \n\nIf you\'re using a hardware RNG, then you need to know the probabilities associated with the data it gives you, else you cannot formally find the Shannon entropy (though you could give it a lower bound if you know the probabilities of some, but not all, events). \n\nBut note that in any case, you are dependent on the knowledge of the distribution associated with the byte stream. You can do statistical tests, like you mention, to verify that the output "looks random" (from a certain perspective). But you\'ll never be able to say any more than "it looks pretty uniformly distributed to me!". You\'ll never be able to look at a bitstream without knowing the distribution and say "there are X bits of entropy here."\n',
]
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

• Datasets: dim_768, dim_512, dim_256, dim_128 and dim_64
• Evaluated with InformationRetrievalEvaluator

Metric	dim_768	dim_512	dim_256	dim_128	dim_64
cosine_accuracy@1	0.4595	0.4538	0.437	0.4098	0.3445
cosine_accuracy@3	0.623	0.6178	0.599	0.5605	0.4933
cosine_accuracy@5	0.682	0.6787	0.6541	0.6138	0.5508
cosine_accuracy@10	0.7415	0.74	0.719	0.6861	0.617
cosine_precision@1	0.4595	0.4538	0.437	0.4098	0.3445
cosine_precision@3	0.2077	0.2059	0.1997	0.1868	0.1644
cosine_precision@5	0.1364	0.1357	0.1308	0.1228	0.1102
cosine_precision@10	0.0742	0.074	0.0719	0.0686	0.0617
cosine_recall@1	0.4595	0.4538	0.437	0.4098	0.3445
cosine_recall@3	0.623	0.6178	0.599	0.5605	0.4933
cosine_recall@5	0.682	0.6787	0.6541	0.6138	0.5508
cosine_recall@10	0.7415	0.74	0.719	0.6861	0.617
cosine_ndcg@10	0.6	0.5958	0.5766	0.5443	0.4765
cosine_mrr@10	0.5547	0.5498	0.5312	0.4994	0.432
cosine_map@100	0.5602	0.555	0.537	0.5058	0.4398

Training Details

Training Dataset

json

• Dataset: json
• Size: 35,129 training samples
• Columns: anchor and positive
• Approximate statistics based on the first 1000 samples:

  	anchor	positive
  type	string	string
  details	min: 16 tokens mean: 182.52 tokens max: 512 tokens	min: 14 tokens mean: 321.62 tokens max: 512 tokens

• Samples:

  anchor	positive
  Are there any common expectations from perspective employers when they hire a Perl developer? For a student who likes Perl and Linux and would like to get a job as Perl developer, what would you recommend to learn? I am looking for things that are generic and applicable to most/all Perl positions, as opposed to specific details of a given company's requirements. In other words, what are the things I should be able to to/know to become more attractive to ANY company looking for a Perl developer.	Some points: As a Perl developer, pretty much any company will expect you to know MORE than Perl. Even in pure Perl shop, you need to know (ideally) JavaScript/overall web development; and SQL for back-end work. And most companies have a mix of languages, so you should be prepared to be Perl/C++ or Perl/Java or whatever else is needed. Much as the fact grates on me, there aren't all that many good "Perl-only" shops I'm aware of. As with any language, a company would expect you to use the language effectively. This has several facets, some are more important in Perl Available libraries. This is a MAJOR point for Perl, of course. Great familiarity with CPAN and knowing which libraries are considered "state of the art"/"most common" for specific common tasks is a must. Can you rattle off - without asking SO - the "standard" library for loading a CSV file? For parsing data out of HTML document? For writing unit tests? For mocking objects? For generating JSON data? For reading simple ...
  I have performed a repeated measures ANOVA in R, as follows: aov_velocity = aov(Velocity ~ Material + Error(Subject/(Material)), data=scrd) summary(aov_velocity) What syntax in R can be used to perform a post hoc test after an ANOVA with repeated measures? Would Tukey's test with Bonferroni correction be appropriate? If so, how could this be done in R?	What you could do is specify the model with lme and then use glht from the multcomp package to do what you want. However, lme gives slightly different F-values than a standard ANOVA (see also my recent questions here). lme_velocity = lme(Velocity ~ Material, data=scrd, random = ~1
  Why are solderless protoboards called "breadboards"? I've used the term for decades but couldn't answer a student's question about the name.	This terminology goes waaaaay back to the days of vacuum tubes. Generally, you would mount a number of tube-sockets on standoffs to a piece of wood (the actual "breadboard"), and do all the wiring with point-point wire and the components just hanging between the various devices. If you needed additional connection points, you would use a solder-lug terminal strip. Image credit: Random googling. The story goes that an engineer had an idea for a vacuum tube device late one night. Looking around the house, the only base for his prototype that he found was indeed his wife's breadboard, from the breadbox. Now, I'm not endorsing actually using a real breadboard. It's your marital strife if you do. I've actually constructed a tube project using the breadboard technique. It works very well.

• Loss: MatryoshkaLoss with these parameters:

  {
      "loss": "MultipleNegativesRankingLoss",
      "matryoshka_dims": [
          768,
          512,
          256,
          128,
          64
      ],
      "matryoshka_weights": [
          1,
          1,
          1,
          1,
          1
      ],
      "n_dims_per_step": -1
  }
  

Training Hyperparameters

Non-Default Hyperparameters

• eval_strategy: epoch
• per_device_train_batch_size: 32
• per_device_eval_batch_size: 16
• gradient_accumulation_steps: 16
• learning_rate: 2e-05
• num_train_epochs: 4
• lr_scheduler_type: cosine
• warmup_ratio: 0.1
• bf16: True
• tf32: True
• load_best_model_at_end: True
• optim: adamw_torch_fused
• batch_sampler: no_duplicates

All Hyperparameters

Click to expand

• overwrite_output_dir: False
• do_predict: False
• eval_strategy: epoch
• prediction_loss_only: True
• per_device_train_batch_size: 32
• per_device_eval_batch_size: 16
• per_gpu_train_batch_size: None
• per_gpu_eval_batch_size: None
• gradient_accumulation_steps: 16
• 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: 4
• max_steps: -1
• lr_scheduler_type: cosine
• 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: True
• 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_fused
• 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: no_duplicates
• multi_dataset_batch_sampler: proportional

Training Logs

Epoch	Step	Training Loss	dim_768_cosine_ndcg@10	dim_512_cosine_ndcg@10	dim_256_cosine_ndcg@10	dim_128_cosine_ndcg@10	dim_64_cosine_ndcg@10
0.1457	10	52.8168	-	-	-	-	-
0.2914	20	41.1964	-	-	-	-	-
0.4372	30	35.2022	-	-	-	-	-
0.5829	40	34.046	-	-	-	-	-
0.7286	50	32.9287	-	-	-	-	-
0.8743	60	29.2281	-	-	-	-	-
0.9909	68	-	0.5965	0.5903	0.5725	0.5365	0.4658
1.0291	70	32.0647	-	-	-	-	-
1.1749	80	26.1728	-	-	-	-	-
1.3206	90	26.8629	-	-	-	-	-
1.4663	100	25.7405	-	-	-	-	-
1.6120	110	25.7913	-	-	-	-	-
1.7577	120	25.4369	-	-	-	-	-
1.9035	130	25.8973	-	-	-	-	-
1.9909	136	-	0.5996	0.5952	0.5750	0.5424	0.4737
2.0583	140	24.895	-	-	-	-	-
2.2040	150	22.9015	-	-	-	-	-
2.3497	160	21.4723	-	-	-	-	-
2.4954	170	22.3792	-	-	-	-	-
2.6412	180	21.1436	-	-	-	-	-
2.7869	190	24.041	-	-	-	-	-
2.9326	200	22.6244	-	-	-	-	-
2.9909	204	-	0.5999	0.5958	0.5761	0.5442	0.4756
3.0874	210	24.6957	-	-	-	-	-
3.2332	220	21.462	-	-	-	-	-
3.3789	230	20.3116	-	-	-	-	-
3.5246	240	20.362	-	-	-	-	-
3.6703	250	21.9136	-	-	-	-	-
3.8160	260	22.2618	-	-	-	-	-
3.9617	270	20.3973	-	-	-	-	-
3.9909	272	-	0.6	0.5958	0.5766	0.5443	0.4765

• The bold row denotes the saved checkpoint.

Framework Versions

• Python: 3.12.8
• Sentence Transformers: 3.4.1
• Transformers: 4.51.3
• PyTorch: 2.7.0+cu126
• 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",
}

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}
}