`sentence_transformers.datasets` contains classes to organize your training input examples.
## ParallelSentencesDataset
`ParallelSentencesDataset` is used for multilingual training. For details, see [multilingual training](../../examples/training/multilingual/README.md).
`sentence_transformers.losses` defines different loss functions that can be used to fine-tune embedding models on training data. The choice of loss function plays a critical role when fine-tuning the model. It determines how well our embedding model will work for the specific downstream task.
Sadly, there is no "one size fits all" loss function. Which loss function is suitable depends on the available training data and on the target task. Consider checking out the [Loss Overview](../../sentence_transformer/loss_overview.html) to help narrow down your choice of loss function(s).
For each sentence pair, we pass sentence A and sentence B through our network which yields the embeddings *u* und *v*. The similarity of these embeddings is computed using cosine similarity and the result is compared to the gold similarity score.
This allows our network to be fine-tuned to recognize the similarity of sentences.
*MultipleNegativesRankingLoss* is a great loss function if you only have positive pairs, for example, only pairs of similar texts like pairs of paraphrases, pairs of duplicate questions, pairs of (query, response), or pairs of (source_language, target_language).
`sentence_transformers.models` defines different building blocks, that can be used to create SentenceTransformer networks from scratch. For more details, see [Training Overview](../training/overview.md).
`sentence_transformers.quantization` defines different helpful functions to perform embedding quantization.
```eval_rst
.. note::
`Embedding Quantization <../../../examples/applications/embedding-quantization/README.html>`_ differs from model quantization. The former shrinks the size of embeddings such that semantic search/retrieval is faster and requires less memory and disk space. The latter refers to lowering the precision of the model weights to speed up inference. This page only shows documentation for the former.
@@ -6,7 +6,6 @@ The training data constist of over 500k examples, while the complete corpus con
## Version Histroy
As we work on the topic, we will publish updated (and improved) models.
### v1
Version 1 models were trained on the training set of MS Marco Passage retrieval task. The models were trained using in-batch negative sampling via the MultipleNegativesRankingLoss with a scaling factor of 20 and a batch size of 128.
Loss functions play a critical role in the performance of your fine-tuned model. Sadly, there is no "one size fits all" loss function. Ideally, this table should help narrow down your choice of loss function(s) by matching them to your data formats.
```eval_rst
.. note::
You can often convert one training data format into another, allowing more loss functions to be viable for your scenario. For example, ``(sentence_A, sentence_B) pairs`` with ``class`` labels can be converted into ``(anchor, positive, negative) triplets`` by sampling sentences with the same or different classes.
These loss functions can be seen as *loss modifiers*: they work on top of standard loss functions, but apply those loss functions in different ways to try and instil useful properties into the trained embedding model.
For example, models trained with <ahref="../package_reference/sentence_transformer/losses.html#matryoshkaloss">`MatryoshkaLoss`</a> produce embeddings whose size can be truncated without notable losses in performance, and models trained with <ahref="../package_reference/sentence_transformer/losses.html#adaptivelayerloss">`AdaptiveLayerLoss`</a> still perform well when you remove model layers for faster inference.
In practice, not all loss functions get used equally often. The most common scenarios are:
*`(anchor, positive) pairs` without any labels: <ahref="../package_reference/sentence_transformer/losses.html#multiplenegativesrankingloss"><code>MultipleNegativesRankingLoss</code></a> is commonly used to train the top performing embedding models. This data is often relatively cheap to obtain, and the models are generally very performant. <ahref="../package_reference/sentence_transformer/losses.html#cachedmultiplenegativesrankingloss"><code>CachedMultipleNegativesRankingLoss</code></a> is often used to increase the batch size, resulting in superior performance.
*`(sentence_A, sentence_B) pairs` with a `float similarity score`: <ahref="../package_reference/sentence_transformer/losses.html#cosinesimilarityloss"><code>CosineSimilarityLoss</code></a> is traditionally used a lot, though more recently <ahref="../package_reference/sentence_transformer/losses.html#cosentloss"><code>CoSENTLoss</code></a> and <ahref="../package_reference/sentence_transformer/losses.html#angleloss"><code>AnglELoss</code></a> are used as drop-in replacements with superior performance.
## Custom Loss Functions
```eval_rst
Advanced users can create and train with their own loss functions. Custom loss functions only have a few requirements:
- They must be a subclass of :class:`torch.nn.Module`.
- They must have ``model`` as the first argument in the constructor.
- They must implement a ``forward`` method that accepts ``sentence_features`` and ``labels``. The former is a list of tokenized batches, one element for each column. These tokenized batches can be fed directly to the ``model`` being trained to produce embeddings. The latter is an optional tensor of labels. The method must return a single loss value.
To get full support with the automatic model card generation, you may also wish to implement:
- a ``get_config_dict`` method that returns a dictionary of loss parameters.
- a ``citation`` property so your work gets cited in all models that train with the loss.
We provide various pre-trained Sentence Transformers models via our Sentence Transformers Hugging Face organization. Additionally, over 6,000 community Sentence Transformers models have been publicly released on the Hugging Face Hub. All models can be found here:
***Original models**: [Sentence Transformers Hugging Face organization](https://huggingface.co/models?library=sentence-transformers&author=sentence-transformers).
***Community models**: [All Sentence Transformer models on Hugging Face](https://huggingface.co/models?library=sentence-transformers).
Each of these models can be easily downloaded and used like so:
```eval_rst
.. sidebar:: Original Models
For the original models from the `Sentence Transformers Hugging Face organization <https://huggingface.co/models?library=sentence-transformers&author=sentence-transformers>`_, it is not necessary to include the model author or organization prefix. For example, this snippet loads `sentence-transformers/all-mpnet-base-v2 <https://huggingface.co/sentence-transformers/all-mpnet-base-v2>`_.
Consider using the `Massive Textual Embedding Benchmark leaderboard <https://huggingface.co/spaces/mteb/leaderboard>`_ as an inspiration of strong Sentence Transformer models. Be wary:
- **Model sizes**: it is recommended to filter away the large models that might not be feasible without excessive hardware.
- **Experimentation is key**: models that perform well on the leaderboard do not necessarily do well on your tasks, it is **crucial** to experiment with various promising models.
```
## Original Models
The following table provides an overview of a selection of our models. They have been extensively evaluated for their quality to embedded sentences (Performance Sentence Embeddings) and to embedded search queries & paragraphs (Performance Semantic Search).
The **all-*** models were trained on all available training data (more than 1 billion training pairs) and are designed as **general purpose** models. The [**all-mpnet-base-v2**](https://huggingface.co/sentence-transformers/all-mpnet-base-v2) model provides the best quality, while [**all-MiniLM-L6-v2**](https://huggingface.co/sentence-transformers/all-MiniLM-L6-v2) is 5 times faster and still offers good quality. Toggle *All models* to see all evaluated original models.
The following models have been specifically trained for **Semantic Search**: Given a question / search query, these models are able to find relevant text passages. For more details, see [Usage > Semantic Search](../../examples/applications/semantic-search/README.md).
The following models have been trained on [215M question-answer pairs](https://huggingface.co/sentence-transformers/multi-qa-MiniLM-L6-dot-v1#training) from various sources and domains, including StackExchange, Yahoo Answers, Google & Bing search queries and many more. These model perform well across many search tasks and domains.
These models were tuned to be used with the dot-product similarity score:
| Model | Performance Semantic Search (6 Datasets) | Queries (GPU / CPU) per sec. |
These models produce normalized vectors of length 1, which can be used with dot-product, cosine-similarity and Euclidean distance as the similarity functions:
| Model | Performance Semantic Search (6 Datasets) | Queries (GPU / CPU) per sec. |
The following models have been trained on the [MSMARCO Passage Ranking Dataset](https://github.com/microsoft/MSMARCO-Passage-Ranking), which contains 500k real queries from Bing search together with the relevant passages from various web sources. Given the diversity of the MSMARCO dataset, models also perform well on other domains.
These models were tuned to be used with the dot-product similarity score:
| Model | MSMARCO MRR@10 dev set | Performance Semantic Search (6 Datasets) | Queries (GPU / CPU) per sec. |
These models produce normalized vectors of length 1, which can be used with dot-product, cosine-similarity and Euclidean distance as the similarity functions:
| Model | MSMARCO MRR@10 dev set | Performance Semantic Search (6 Datasets) | Queries (GPU / CPU) per sec. |
[MSMARCO Models - More details](../pretrained-models/msmarco-v5.md)
---
## Multilingual Models
The following models similar embeddings for the same texts in different languages. You do not need to specify the input language. Details are in our publication [Making Monolingual Sentence Embeddings Multilingual using Knowledge Distillation](https://arxiv.org/abs/2004.09813). We used the following 50+ languages: ar, bg, ca, cs, da, de, el, en, es, et, fa, fi, fr, fr-ca, gl, gu, he, hi, hr, hu, hy, id, it, ja, ka, ko, ku, lt, lv, mk, mn, mr, ms, my, nb, nl, pl, pt, pt-br, ro, ru, sk, sl, sq, sr, sv, th, tr, uk, ur, vi, zh-cn, zh-tw.
### Semantic Similarity Models
These models find semantically similar sentences within one language or across languages:
-**[distiluse-base-multilingual-cased-v2](https://huggingface.co/sentence-transformers/distiluse-base-multilingual-cased-v2)**: Multilingual knowledge distilled version of [multilingual Universal Sentence Encoder](https://arxiv.org/abs/1907.04307). This version supports 50+ languages, but performs a bit weaker than the v1 model.
-**[paraphrase-multilingual-MiniLM-L12-v2](https://huggingface.co/sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2)** - Multilingual version of [paraphrase-MiniLM-L12-v2](https://huggingface.co/sentence-transformers/paraphrase-MiniLM-L12-v2), trained on parallel data for 50+ languages.
-**[paraphrase-multilingual-mpnet-base-v2](https://huggingface.co/sentence-transformers/paraphrase-multilingual-mpnet-base-v2)** - Multilingual version of [paraphrase-mpnet-base-v2](https://huggingface.co/sentence-transformers/paraphrase-mpnet-base-v2), trained on parallel data for 50+ languages.
### Bitext Mining
Bitext mining describes the process of finding translated sentence pairs in two languages. If this is your use-case, the following model gives the best performance:
-**[LaBSE](https://huggingface.co/sentence-transformers/LaBSE)** - [LaBSE](https://arxiv.org/abs/2007.01852) Model. Supports 109 languages. Works well for finding translation pairs in multiple languages. As detailed [here](https://arxiv.org/abs/2004.09813), LaBSE works less well for assessing the similarity of sentence pairs that are not translations of each other.
Extending a model to new languages is easy by following [Training Examples > Multilingual Models](../../examples/training/multilingual/README.html).
## Image & Text-Models
The following models can embed images and text into a joint vector space. See [Usage > Image Search](../../examples/applications/image-search/README.md) for more details how to use for text2image-search, image2image-search, image clustering, and zero-shot image classification.
The following models are available with their respective Top 1 accuracy on zero-shot ImageNet validation dataset.
We further provide this multilingual text-image model:
-**[clip-ViT-B-32-multilingual-v1](https://huggingface.co/sentence-transformers/clip-ViT-B-32-multilingual-v1)** - Multilingual text encoder for the [clip-ViT-B-32](https://huggingface.co/sentence-transformers/clip-ViT-B-32) model using [Multilingual Knowledge Distillation](https://arxiv.org/abs/2004.09813). This model can encode text in 50+ languages to match the image vectors from the [clip-ViT-B-32](https://huggingface.co/sentence-transformers/clip-ViT-B-32) model.
## INSTRUCTOR models
Some INSTRUCTOR models, such as [hkunlp/instructor-large](https://huggingface.co/hkunlp/instructor-large), are natively supported in Sentence Transformers. These models are special, as they are trained with instructions in mind. Notably, the primary difference between normal Sentence Transformer models and Instructor models is that the latter do not include the instructions themselves in the pooling step.
"Represent the Wikipedia question for retrieving supporting documents: "
)
corpus=[
'Yams are perennial herbaceous vines native to Africa, Asia, and the Americas and cultivated for the consumption of their starchy tubers in many temperate and tropical regions. The tubers themselves, also called "yams", come in a variety of forms owing to numerous cultivars and related species.',
"The disparate impact theory is especially controversial under the Fair Housing Act because the Act regulates many activities relating to housing, insurance, and mortgage loans—and some scholars have argued that the theory's use under the Fair Housing Act, combined with extensions of the Community Reinvestment Act, contributed to rise of sub-prime lending and the crash of the U.S. housing market and ensuing global economic recession",
"Disparate impact in United States labor law refers to practices in employment, housing, and other areas that adversely affect one group of people of a protected characteristic more than another, even though rules applied by employers or landlords are formally neutral. Although the protected classes vary by statute, most federal civil rights laws protect based on race, color, religion, national origin, and sex as protected traits, and some laws include disability status and other traits as well.",
]
corpus_instruction="Represent the Wikipedia document for retrieval: "
All other Instructor models either 1) will not load as they refer to `InstructorEmbedding` in their `modules.json` or 2) require calling `model.set_pooling_include_prompt(include_prompt=False)` after loading.
## Scientific Similarity Models
[SPECTER](https://arxiv.org/abs/2004.07180) is a model trained on scientific citations and can be used to estimate the similarity of two publications. We can use it to find similar papers.
