Model description

The project aims to train sentence embedding models on very large sentence level datasets using a self-supervised contrastive learning objective. We used the pretrained 'MiniLM-L12' model and fine-tuned in on a 1B sentence pairs dataset. We use a contrastive learning objective: given a sentence from the pair, the model should predict which out of a set of randomly sampled other sentences, was actually paired with it in our dataset.

We developped this model during the Community week using JAX/Flax for NLP & CV, organized by Hugging Face. We developped this model as part of the project: Train the Best Sentence Embedding Model Ever with 1B Training Pairs. We benefited from efficient hardware infrastructure to run the project: 7 TPUs v3-8, as well as intervention from Google’s Flax, JAX, and Cloud team member about efficient deep learning frameworks.

Intended uses

Our model is intented to be used as a sentence encoder. Given an input sentence, it ouptuts a vector which captures the sentence semantic information. The sentence vector may be used for information retrieval, clustering or sentence similarity tasks.

How to use

Here is how to use this model to get the features of a given text using SentenceTransformers library:

from sentence_transformers import SentenceTransformer

model = SentenceTransformer('flax-sentence-embeddings/all_datasets_v4_MiniLM-L12')
text = "Replace me by any text you'd like."
text_embbedding = model.encode(text)
# array([-0.01559514,  0.04046123,  0.1317083 ,  0.00085931,  0.04585106,
#        -0.05607086,  0.0138078 ,  0.03569756,  0.01420381,  0.04266302 ...],
#        dtype=float32)

Training procedure

Pre-training

We use the pretrained 'MiniLM-L12'. Please refer to the model card for more detailed information about the pre-training procedure.

Fine-tuning

We fine-tune the model using a contrastive objective. Formally, we compute the cosine similarity from each possible sentence pairs from the batch. We then apply the cross entropy loss by comparing with true pairs.

Hyper parameters

We trained ou model on a TPU v3-8. We train the model during 540k steps using a batch size of 1024 (128 per TPU core). We use a learning rate warm up of 500. The sequence length was limited to 128 tokens. We used the AdamW optimizer with a 2e-5 learning rate. The full training script is accessible in this current repository.

Training data

We use the concatenation from multiple datasets to fine-tune our model. The total number of sentence pairs is above 1 billion sentences. We sampled each dataset given a weighted probability which configuration is detailed in the data_config.json file.

Dataset	Paper	Number of training tuples
GOOAQ: Open Question Answering with Diverse Answer Types	paper	3,012,496
Stack Exchange	-	364,001
Flickr 30k	paper	317,695
[COCO 2020](COCO 2020)	paper	828,395
Code Search	-	1,151,414
TriviaqQA	-	73,346
SQuAD2.0	paper	87,599
Natural Questions (NQ)	paper	100,231
Simple Wikipedia	paper	102,225
Quora Question Pairs	-	103,663
Altlex	paper	112,696
Wikihow	paper	128,542
Sentence Compression	paper	180,000
AllNLI (SNLI and MultiNLI	paper SNLI, paper MultiNLI	277,230
Eli5	paper	325,475
SPECTER	paper	684,100
S2ORC Title/Abstract	paper	41,769,185
S2ORC Citation/Citation	paper	52,603,982
S2ORC Citation/Abstract	paper	116,288,806
PAQ	paper	64,371,441
WikiAnswers	paper	77,427,422
SearchQA	-	582,261
Yahoo Answers Title/Answer	paper	1,198,260
Yahoo Answers Title/Question	paper	659,896
Yahoo Answers Question/Answer	paper	681,164
MS MARCO	paper	9,144,553
Reddit conversationnal	paper	726,484,430
total		1,097,953,922