m7n/bge-m3-philosophy-triplets_v3

SentenceTransformer based on BAAI/bge-m3

This is a sentence-transformers model finetuned from BAAI/bge-m3. It maps sentences & paragraphs to a 1024-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-m3
• Maximum Sequence Length: 8192 tokens
• Output Dimensionality: 1024 tokens
• 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': 8192, 'do_lower_case': False}) with Transformer model: XLMRobertaModel 
  (1): Pooling({'word_embedding_dimension': 1024, '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("m7n/bge-m3-philosophy-triplets_v3")
# Run inference
sentences = [
    'met, Peter van Inwagen, a libertarian, concludes that free will is a mystery. Recently, the Mind Argument has drawn a number of criticisms. Here I seek to add to its woes. Quite apart from its other problems, I argue, the Mind Argument does a poor job of isolating the important concern for libertarians that it raises. Once this concern has been clarified,',
    'however, another argument serves to renew the challenge. The Assimilation Argument challenges libertarians to explain how ostensible exercises of free will are relevantly different from other causally undetermined outcomes, outcomes that nobody would count as exercises of free will. In particular, libertarians must explain how agents can have the power',
    'contended that the Assimilation Argument is unsound. Here I defend the Assimilation Argument and the Rollback Argument, a second challenge to libertarianism that Franklin rejects. My aim in doing so is to underscore the force of these challenges, and thereby to resist what appears to be an emerging trend in',
]
embeddings = model.encode(sentences)
print(embeddings.shape)
# [3, 1024]

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

Evaluation

Metrics

Triplet

• Dataset: m3
• Evaluated with TripletEvaluator

Metric	Value
cosine_accuracy	0.978
dot_accuracy	0.022
manhattan_accuracy	0.974
euclidean_accuracy	0.978
max_accuracy	0.978

Triplet

• Dataset: all-nli-test
• Evaluated with TripletEvaluator

Metric	Value
cosine_accuracy	0.9755
dot_accuracy	0.0245
manhattan_accuracy	0.976
euclidean_accuracy	0.9755
max_accuracy	0.976

Training Details

Training Dataset

Unnamed Dataset

• Size: 5,000 training samples
• Columns: anchor, positive, and negative
• Approximate statistics based on the first 1000 samples:

  	anchor	positive	negative
  type	string	string	string
  details	min: 17 tokens mean: 288.26 tokens max: 566 tokens	min: 21 tokens mean: 288.3 tokens max: 565 tokens	min: 22 tokens mean: 301.96 tokens max: 577 tokens

• Samples:

  anchor	positive	negative
  Reform (Boston: Auburn House, 1982). 7 See, e.g., K. Polk, 'Rape Reform and Criminal Justice Proces;ing', 31 Crime & Delinquency 191 (1985); W. Loh, 'The Impact of Common Law and Reform Rape Statutes on Prosecution: An Empirical Study', 55 Wash. L. Rev. 543, 552-54 (1980); J. Marsh, A. Geist & N. Caplan, Rape and the Limits of Law Reform 65 (1982). 8 See S. Estrich, Real Rape (Cambridge: Harvard University Press, 1987). 7 Susan Estrich throw their hands up in the air, depending on whether they saw themselves as feminists, or wanted me to. But nearly all of them said the same thing then, and most of them continue to repeat it even today: juries won't convict. You may think it's rape, and I may even think it's rape, but the public doesn't, and we've got too many real rapes in this office to waste our time on cases where the woman will eventually back out, and the jury would never convict. It hasn't been easy to dissuade them. Certainly, what went on in most law school classrooms in the nineteen-eighties, where rape was so politicized that one could hardly debate it, provides no answer to the real world concerns, or at least justifications, that animate so many prosecutors' offices. When I started teaching law in 1981, there was still nothing in the casebooks on rape, so I put my own materials together. Hard cases make good classes. I found better examples than the three English fools,	more sympathetic men: doctors, not drunken sailors. And women who didn't scream quite so much, or fight back quite so vigorously. Two great classes, as these things go. Most of the men I knew were reluctant to teach rape, for reasons that at first I didn't quite understand. But I convinced one to try the next year, and lent him my materials, and watched first with amusement and then with dismay when a delegation of women students came to visit him before class to complain that his (my) materials were sexist and inappropriate. I'd hoped those days were over. After all, all the casebooks now have sections on rape, and everyone covers it somehow. But only last week, a colleague who is one of the best criminal law professors in the country told me of being visited by a similar delegation in 1991, a decade later. They didn't like the way he taught rape. There were victims in the class. They didn't want to hear "the other side". In their view, there was no "other side". Nothing to debate. Lord knows, I'm sympathetic to those who have been victimized and who feel that they have nothing to learn from a debate about rape. Maybe they don't. But the rest of us, the rest of their classmates, the rest of the country, surely do. Silencing the debate only reenforces and reaffirms the status quo that many of us are trying to change. The debate I am interested in is not between the radical feminists	relations, thwarting possibilities of knowing the specific harms particular acts of rape enact well enough to render intelligible their far-reaching social-political-moral implications. Taking my point of departure from Debra Bergoffen's call for 'a new epistemology of rape', I
  framed, because the key premissthe thesis of universal causal determinismis, in the domain of human behaviour, an unjustified conjecture based on over-simplified, under-informed explanatory models. Kant's semantics of singular cognitive reference (explained herein), which stands independently of his Transcendental Idealism, justifies and emphasises a quadruple distinction between causal description, causal ascription (predication), (approximately) true causal ascription (accurate predication) and cognitively justified causal ascription. Contemporary causal theories of mind, of action	or of meaning do not suffice for causal ascription, and so cannot suffice for causal predication, and hence cannot justify causal determinism about human behaviour. More generally, the principle of universal causal determinism is a regulative principle governing causal inquiry and was so formulated by LaPlace. Only successful, sufficient causal explanation of particular events provides for causal knowledge of those events. Such knowledge we lack in the domain of human behaviour. Rational belief, including scientific belief, requires apportioning belief to justifying evidence; all else	etc.; but causal explanation, in its advanced stages, is not content with such vague statements of uniformities. The uniformities which are sought are the ones which can be correlated with precise, numerical relations. A third characteristic which has often been associated with causality is the concept of determinism or necessity. This third trait of causality has been violently criticized. From Hume to Russell it has been a common contention that necessity is a fiction and that causality merely means invariable connection of uniformity of processes. For the present, without analysing the objections raised to the necessity or determinism of causal relations, I shall limit myself to brief comments on the meaning of necessity, and on the specific sense of necessity in the causal relation. Though necessity, or determinism, should be a fiction or a myth, it is certainly not a meaningless concept. Hume, who presents the most effective objections to the concept of necessity, was quite certain that the concept had a meaning, since otherwise his criticisms would have been of no importance. What then is necessity? Briefly, necessity refers to deductive systems where from self-evident, postu' S. Alexander, Space, Time and Deity, vol. i, p. 288. 320 CAUSALITY lated, or "convenient" concepts other concepts are derived. The clearest illustration of the notion of necessity is the concept of classes where the implicative relation is that of inclusion, or that of from species to genus. Causal relations are deterministic in the sense that cause and effect contain
  presuppositions. In a nutshell view of aspects of the history of science the presuppositions, as part of the relevant historical context, operate within the scientific enterprise. Science used to be views as positivist and final; yet, the logic of methodology, from E.A. Burtt to T.S. Kuhn and Karl Popper	show that presuppositions are part of the scientific research. Particularly the pioneering work of the archeologist and philosopher R.G. Collingwood and the seminal thinker S.N. Nasr, who opened the doors for the inquiry of the role of science within Islamic Civilization, are highlighted. We show that the	continues to evolve, it seems most probable that it will someday diverge again from its seeming congruity with the teachings of the Qur'an. An even sharper critique arises from the unexpressed as sumption that there exists a standard of knowledge on a par with, if not actually higher than, Allah's revealed truth to which that revelation must be submitted for its validation. Thus, while this attempt to claim unity (or at least non-contradiction) between modern science and Islamic revelation does not require any sort of limitation on science or scientific research, it remains unacceptable to the majority of modern scientists, as well as to many Muslim thinkers, because of the implications of its assumptions.8 The first assumption, that science has reached its peak of development, is unacceptable to science. The second assumption, that religious dogma should, or even can, be judged by the tenets of modern science, is rejected by many Muslim thinkers. Scientific Knowledge Needs an Islamic Interpretation In response to the unacceptable and often Positivistic analysis of Bucaille, the Islamic intellectual community has adopted more radical al ternatives, such as Ismail al-Faruqi's call for an Islamization of all knowledge, including modern scientific knowledge.9 His argument grows out of a perception that Islamic culture, which once led the Mediterranean world, is increasingly threatened by secularized and individualistic West ern values?a threat that is as much cultural as it is political. Faruqi's plea for an Islamization of knowledge, then, should be read in the context

• Loss: TripletLoss with these parameters:

  {
      "distance_metric": "TripletDistanceMetric.COSINE",
      "triplet_margin": 0.05
  }
  

Evaluation Dataset

Unnamed Dataset

• Size: 500 evaluation samples
• Columns: anchor, positive, and negative
• Approximate statistics based on the first 1000 samples:

  	anchor	positive	negative
  type	string	string	string
  details	min: 29 tokens mean: 288.01 tokens max: 646 tokens	min: 26 tokens mean: 286.17 tokens max: 499 tokens	min: 16 tokens mean: 289.98 tokens max: 544 tokens

• Samples:

  anchor	positive	negative
  public have now before them, the evidences of natural religion, the evidences of revealed religion, and an account of the duties that result from both" (Paley 1829, v-vi ). Without the arguments of natural theology, revelation and religious morality could not be discussed. The evidence of human reason and of the empirical study of the natural world, was the basis for establishing the existence and character of a deity. Only after these were established could one consider other sources of knowledge. Without the arguments available from natural theology that there exists a single unified, beneficent, and omnipresent deity, there is no warrant for accepting that the deity is capable of revelation, that revelation ought to be a trustworthy basis for knowledge, or that there exists a moral order or set of moral duties. The mere existence of a deity was insufficient for Paley. As a work of theology, the most important conclusion is not only that God merely exists, but also that characteristics about God can be known and inferences about the proper relationship of the individual and society to God can be drawn. As a result, much of the early chapters of Natural Theology did not put forth an argument for the existence of God, but instead excluded bad arguments for the existence of God. In order to understand what Paley's argument is, it is paramount to also understand what it is not. 58 ADAM R. SHAPIRO With the recent restoration of Paley's work, several historians and philosophers of science have begun this	re-examination. Sander Gliboff noted that "Paley devoted a chapter to comparing rival hypotheses unfavorably with his own and eliminating them from consideration one by one" (Gliboff 2000). For Gliboff, doing so was essential for establishing the legitimacy of the central argument of the Natural Theology. The central argument, which is what Paley himself referred to as "the argument" in the opening chapters of the book, is an inference that can be formulated as the claim that if an object has a "purpose," then it has a designer. In analyzing this, Gliboff agreed with Eliot Sober's 1993 assessment that Paley's central argument was based upon "inference to the best explanation" (Sober 2000, 31). Sober ascribed to Paley two possible explanations for apparent purposeful objects, design and random chance. The explanation that purposeful objects have their origins in a designer is pitted against the other possible explanation for the appearance of purpose in an object and is ruled the most likely. Gliboff showed that these were not the only contenders and claims that Paley refutes several alternatives that make up a "pool" of possible explanations. The inference to a designer is judged the most likely. If Natural Theology was in fact based upon an inference to the best explanation, then it becomes clear why a Darwinian conclusion might be seen to uproot Paley. If the workings of nature can be seen to generate purpose without the apparent intentional workings of a designer, then alternate explanations are strengthened and	and intellectual sides of the stratagem. In response to Erasmus Darwin's evolutionary theories. Paley asserted that the divinely designed architecture of nature had remained unchanged since the creation. But the more he emphasized the preordained nature of providence, its effectuation through mechanical dispositions, the less room there appeared to be for particular interventions. Section 2 concentrates on Paley's efforts to reconcile this model of a law-governed, mechanical universe, with the belief
  the two lines of work from Sections 2 and 3 together: the realization that there is no in principle difference between functional and natural kinds regarding their epistemic importance for scientific reasoning and explanation provides clues how the problem for HPC theory that I pointed out in Section 2 can be resolved, and HPC theory can be turned into a full‐fledged theory of natural kindhood able to cover kinds in the special sciences as well as the less controversial natural kinds. 1. Brigandt (2009) recently reached a similar conclusion on the basis of different considerations. 2. A Problem for HPC Theory. In the philosophical literature on the topic, two distinct ways of thinking about natural kinds can be found. On the one hand, there is the essentialist tradition that, broadly taken, understands natural kinds as groupings of things according to their natures, their intrinsic properties or causal capacities, their microscopic structures, and so on. On the other hand, there is the more recent tradition that understands natural kinds as groupings of things over which we can make reliable inductions. That these lines of work really are quite distinct can be seen from the way in which they conceive of the problem of natural kinds.2 The former line of work conceives of the problem as a metaphysical problem, that is, as the question, What sorts of things are there in the world? As Brian Ellis put it in a recent defense of essentialism, “membership of a natural kind is	decided by nature, not by us” (2001, 19). The latter line of work, in contrast, sees it as foremost a question in epistemology, that is, as the question, Which ways of grouping things are best suited to help us make inferences and to explain phenomena? Boyd, for example, asserted that “it is a truism that the philosophical theory of natural kinds is about how classificatory schemes come to contribute to the epistemic reliability of inductive and explanatory practices” (1999a, 146; also 1999b, 69). On this view, kind membership is decided more by us than by nature. With respect to the explanation‐grounding capacity of natural kinds, the two lines of work provide different answers that run into different kinds of problems. From the perspective of the essentialist tradition it should be no miracle that natural kinds ground reliable inferences and explanations. If there is a definitive, objective way in which the world is made up of kinds of things, then clearly any explanation of a given phenomenon should ultimately make reference to some of these objectively existing kinds of things. Once we have achieved an inventory of the various kinds that exist in the world and have an account of their metaphysics, we have a theory of the kinds that can feature in our explanations. The problem, however, is that we do not have any direct access to the world that would allow us to compile the required inventory of the world’s furniture. Our best bet at obtaining such an inventory is to	There is, however, a selection of strong papers in the collection that do address key topics in the contemporary debate concerning natural kinds— too many, in fact, to review them all here (and hope to do them justice). As such, I shall mention a selection that struck me as of particular interest, since each addresses one of the perennial themes in the natural kinds debate, namely the alleged epistemic utility of natural kinds, and their role in explanatory and inductive reasoning. I'll start with Chapter 2 of the collection, a paper written by Peter Godfrey-Smith. In this chapter, Godfrey-Smith discusses inductive inferences and their associated justification. Inductive inferences are a perennial philo sophical topic, cropping up in both the philosophy of science and epistemol ogy. In the context of the topic of the edited collection, there are apparent inductive benefits bestowed by recognizing and classifying objects into natural kinds, which makes induction a central topic to the broader debate concerning natural kinds. Godfrey-Smith identifies what he takes to be two conflated types of induct ive inference and goes on to sketch a philosophical justification for each, re spectively, which has not been used to justify what philosophers have historically referred to as 'inductive' inferences. The approach employed builds upon Frank Jackson's response to Nelson Goodman's classic 'grue' example, from his new riddle of induction. Godfrey-Smith sketches a method of inference that is built upon statistical models of population sampling, where there are rules telling us how the dis tribution
  model, can be chosen in several ways from the set of all elementary experiences. In case (ii) it is obviously conceivable that the mental stream, like a loaf of bread, can be cut into slices in more than one way. At the same time (in ? 67) that Carnap is inclined to consider elementary experiences as abstractions of some kind (as mere 'places' in the stream), he maintains that they are what is "primarily given" both psychologically and epistemologically. The set of all elementary experiences occurring in one's own mental life is symbolically designated by erl. When I wish to show that the owner is a person A, I shall write erlA. When speaking, in what follows, of elementary experiences I shall always assume that we are dealing with experiences of one and the same person unless the contrary is explicitly stated. 1.3. Quality Points The visual field that is a part of my present elementary experience is assumed in the Aufbau to be a two-dimensional spatial configuration of color points. Each color point is characterized by its two 'local signs', which together determine its place in the visual field, and by its hue, saturation, and brightness, which together determine its place in the color solid.9 In this sense, the color points occurring in the elementary expe riences can be thought of as points in a five-dimensional space (the word 'space' here understood in an abstract sense), which may be called the HOW CARNAP BUILT THE WORLD IN	1928 341 color space.10 The color points occurring in a given elementary experience will then be a certain subset of the color space. In that space any two points whatsoever have a mutual distance. (How the distance is to be measured is left unexplained in the Aufbau.) If the distance is less than or equal to a certain standard distance, the two points are said to be similar (?hnlich).11 This notion of similarity thus involves the notion of a standard distance. Carnap seems to consider this distance as small, but otherwise he says nothing about it. (The question as to how the choice of standard distance affects the constructions of the Aufbau is disturbing and impossible to answer.) If the points p and p' are similar and if p occurs in the elementary experience e and/?' in e9 then e and e' are said to be part similar (teil?hnlich).,12 Color points are a species of the genus quality points. On one level of thought, in one sense, Carnap considers an elementary experience as a complex of a number of quality points of various kinds. One kind are visual color points, another auditive qualities, a third olfactory, etc. He is also prepared to recognize emotive, volitive, and intellectual qualities (quality points). The word 'sense class' is used to denote such a kind of quality.13 For each sense class, just as for vision, notions of distance and standard distance are postulated. Two points from the same sense class are said to be similar if	presentational of worldly properties, but as blank sensations, extrinsically endowed with representational powers in virtue of standing as 'natural signs' for their normal causes.1 On that view, both colour perceptions and colour beliefs normally correspond with reality, in the simple sense of being true. Still, the view retains a strong case to be regarded a form of anti-realism. This has nothing to do with a failure of correspondence in the sense of plain truth; nor is it to be characterized as a failure of any other kind of 'correspondence'. It is rather to be brought out by means of a contrast in directions of explanation with respect to the individuation of perceptions of colour and the colours of things. Compare the case of shapes. The most basic distinctions are between squareness and circularity, say, as properties of things in the world. Having first identified which property squareness is, we may then identify perceptions of squareness as those which present something as having that property. On the current view concerning colour, the direction of explanation is the reverse. The most basic distinctions are between experiences of redness and experiences of greenness, say, conceived as blank sensations. Having made such distinctions, we may then define a property redness which applies to mind-independent objects, as that of being disposed to produce those experiences red-type ones or as the property of having whatever physical constitution actually grounds that disposition. Thus, although, the colours are perfectly real, in the sense that representations

• Loss: TripletLoss with these parameters:

  {
      "distance_metric": "TripletDistanceMetric.COSINE",
      "triplet_margin": 0.05
  }
  

Training Hyperparameters

Non-Default Hyperparameters

• eval_strategy: steps
• per_device_train_batch_size: 4
• per_device_eval_batch_size: 4
• learning_rate: 1e-05
• warmup_ratio: 0.1
• 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: 4
• per_device_eval_batch_size: 4
• per_gpu_train_batch_size: None
• per_gpu_eval_batch_size: None
• gradient_accumulation_steps: 1
• eval_accumulation_steps: None
• learning_rate: 1e-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: 3
• 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: 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: 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
• eval_on_start: False
• batch_sampler: no_duplicates
• multi_dataset_batch_sampler: proportional

Training Logs

Epoch	Step	Training Loss	loss	all-nli-test_max_accuracy	m3_max_accuracy
0	0	-	-	-	0.91
0.08	100	0.0083	0.0101	-	0.916
0.16	200	0.0094	0.0084	-	0.932
0.24	300	0.0075	0.0070	-	0.942
0.32	400	0.0085	0.0065	-	0.952
0.4	500	0.0068	0.0058	-	0.956
0.48	600	0.0064	0.0057	-	0.958
0.56	700	0.0063	0.0051	-	0.964
0.64	800	0.0064	0.0047	-	0.974
0.72	900	0.0049	0.0044	-	0.974
0.8	1000	0.0057	0.0043	-	0.966
0.88	1100	0.0042	0.0054	-	0.96
0.96	1200	0.0037	0.0045	-	0.972
1.04	1300	0.006	0.0056	-	0.962
1.12	1400	0.0043	0.0053	-	0.96
1.2	1500	0.0026	0.0045	-	0.966
1.28	1600	0.001	0.0046	-	0.968
1.3600	1700	0.0012	0.0045	-	0.962
1.44	1800	0.0007	0.0042	-	0.968
1.52	1900	0.0006	0.0044	-	0.968
1.6	2000	0.0013	0.0040	-	0.97
1.6800	2100	0.0006	0.0038	-	0.974
1.76	2200	0.0011	0.0036	-	0.97
1.8400	2300	0.0012	0.0036	-	0.966
1.92	2400	0.001	0.0038	-	0.968
2.0	2500	0.0009	0.0038	-	0.974
2.08	2600	0.0006	0.0037	-	0.976
2.16	2700	0.0006	0.0037	-	0.98
2.24	2800	0.0004	0.0035	-	0.978
2.32	2900	0.0003	0.0034	-	0.978
2.4	3000	0.0003	0.0034	-	0.978
2.48	3100	0.0001	0.0033	-	0.978
2.56	3200	0.0003	0.0034	-	0.978
2.64	3300	0.0003	0.0034	-	0.976
2.7200	3400	0.0001	0.0033	-	0.976
2.8	3500	0.0003	0.0033	-	0.976
2.88	3600	0.0003	0.0033	-	0.976
2.96	3700	0.0003	0.0033	-	0.978
3.0	3750	-	-	0.976	-

Framework Versions

• Python: 3.10.12
• Sentence Transformers: 3.0.1
• Transformers: 4.42.4
• PyTorch: 2.3.1+cu121
• Accelerate: 0.32.1
• Datasets: 2.21.0
• 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",
}

TripletLoss

@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}
}