multi-qa-MiniLM-L6-cos-v1-GGML

This is a sentence-transformers model aimed to be used with bert.cpp by Gerganov's GGML Library: It maps sentences & paragraphs to a 384 dimensional dense vector space and was designed for semantic search. It has been trained on 215M (question, answer) pairs from diverse sources. For an introduction to semantic search, have a look at: SBERT.net - Semantic Search

Usage (Start Server)

Using this model becomes easy when you have bert.cpp installed:

./build/bin/server -m models/all-MiniLM-L6-v2/ggml-model-q4_0.bin --port 8085

# bert_model_load: loading model from 'models/all-MiniLM-L6-v2/ggml-model-q4_0.bin' - please wait ...
# bert_model_load: n_vocab = 30522
# bert_model_load: n_ctx   = 512
# bert_model_load: n_embd  = 384
# bert_model_load: n_intermediate  = 1536
# bert_model_load: n_head  = 12
# bert_model_load: n_layer = 6
# bert_model_load: f16     = 2
# bert_model_load: ggml ctx size =  13.57 MB
# bert_model_load: ............ done
# bert_model_load: model size =    13.55 MB / num tensors = 101
# Server running on port 8085 with 4 threads
# Waiting for a client

Usage (Start Client)

Then you can use the model like this:

python3 examples/sample_client.py 8085
# Loading texts from sample_client_texts.txt...
# Loaded 1738 lines.
# Starting with a test query "Should I get health insurance?"
# Closest texts:
# 1. Will my Medicare premiums be higher because of my higher income?
#  (similarity score: 0.4844)
# 2. Can I sign up for Medicare Part B if I am working and have health insurance through an employer?
#  (similarity score: 0.4575)
# 3. Should I sign up for Medicare Part B if I have Veterans' Benefits?
#  (similarity score: 0.4052)
# Enter a text to find similar texts (enter 'q' to quit): expensive
# Closest texts:
# 1. It is priced at $ 5,995 for an unlimited number of users tapping into the single processor , or $ 195 per user with a minimum of five users .
#  (similarity score: 0.4597)
# 2. The new system costs between $ 1.1 million and $ 22 million , depending on configuration .
#  (similarity score: 0.4547)
# 3. Each hull will cost about $ 1.4 billion , with each fully outfitted submarine costing about $ 2.2 billion , Young said .
#  (similarity score: 0.4078)

Converting models to ggml format

Converting models is similar to llama.cpp. Use models/convert-to-ggml.py to make hf models into either f32 or f16 ggml models. Then use ./build/bin/quantize to turn those into Q4_0, 4bit per weight models.

There is also models/run_conversions.sh which creates all 4 versions (f32, f16, Q4_0, Q4_1) at once.

cd models
# Clone a model from hf
git clone https://huggingface.co/sentence-transformers/multi-qa-MiniLM-L6-cos-v1
# Run conversions to 4 ggml formats (f32, f16, Q4_0, Q4_1)
sh run_conversions.sh multi-qa-MiniLM-L6-cos-v1

Technical Details

In the following some technical details how this model must be used:

Setting	Value
Dimensions	384
Produces normalized embeddings	Yes
Pooling-Method	Mean pooling
Suitable score functions	dot-product (util.dot_score), cosine-similarity (util.cos_sim), or euclidean distance

Note: When loaded with sentence-transformers, this model produces normalized embeddings with length 1. In that case, dot-product and cosine-similarity are equivalent. dot-product is preferred as it is faster. Euclidean distance is proportional to dot-product and can also be used.

Benchmarks

Running MTEB (Massive Text Embedding Benchmark) with bert.cpp vs. sbert(cpu mode) gives comparable results between the two, with quantization having minimal effect on accuracy and eval time being similar or better than sbert with batch_size=1 (bert.cpp doesn't support batching).

See benchmarks more info.

all-MiniLM-L6-v2

Data Type	STSBenchmark	eval time	EmotionClassification	eval time
f32	0.8201	6.83	0.4082	11.34
f16	0.8201	6.17	0.4085	10.28
q4_0	0.8175	5.45	0.3911	10.63
q4_1	0.8223	6.79	0.4027	11.41
sbert	0.8203	2.74	0.4085	5.56
sbert-batchless	0.8203	13.10	0.4085	15.52

all-MiniLM-L12-v2

Data Type	STSBenchmark	eval time	EmotionClassification	eval time
f32	0.8306	13.36	0.4117	21.23
f16	0.8306	11.51	0.4119	20.08
q4_0	0.8310	11.27	0.4183	20.81
q4_1	0.8325	12.37	0.4093	19.38
sbert	0.8309	5.11	0.4117	8.93
sbert-batchless	0.8309	22.81	0.4117	28.04

bert-base-uncased

bert-base-uncased is not a very good sentence embeddings model, but it's here to show that bert.cpp correctly runs models that are not from SentenceTransformers. Technically any hf model with architecture BertModel or BertForMaskedLM should work.

Data Type	STSBenchmark	eval time	EmotionClassification	eval time
f32	0.4738	52.38	0.3361	88.56
f16	0.4739	33.24	0.3361	55.86
q4_0	0.4940	33.93	0.3375	57.82
q4_1	0.4612	36.86	0.3318	59.63
sbert	0.4729	16.97	0.3527	28.77
sbert-batchless	0.4729	69.97	0.3526	79.02

---

Background

The project aims to train sentence embedding models on very large sentence level datasets using a self-supervised contrastive learning objective. 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 Googles Flax, JAX, and Cloud team member about efficient deep learning frameworks.

Intended uses

Our model is intented to be used for semantic search: It encodes queries / questions and text paragraphs in a dense vector space. It finds relevant documents for the given passages.

Note that there is a limit of 512 word pieces: Text longer than that will be truncated. Further note that the model was just trained on input text up to 250 word pieces. It might not work well for longer text.

Training procedure

The full training script is accessible in this current repository: train_script.py.

Pre-training

We use the pretrained nreimers/MiniLM-L6-H384-uncased model. Please refer to the model card for more detailed information about the pre-training procedure.

Training

We use the concatenation from multiple datasets to fine-tune our model. In total we have about 215M (question, answer) pairs. We sampled each dataset given a weighted probability which configuration is detailed in the data_config.json file.

The model was trained with MultipleNegativesRankingLoss using Mean-pooling, cosine-similarity as similarity function, and a scale of 20.

Dataset	Number of training tuples
WikiAnswers Duplicate question pairs from WikiAnswers	77,427,422
PAQ Automatically generated (Question, Paragraph) pairs for each paragraph in Wikipedia	64,371,441
Stack Exchange (Title, Body) pairs from all StackExchanges	25,316,456
Stack Exchange (Title, Answer) pairs from all StackExchanges	21,396,559
MS MARCO Triplets (query, answer, hard_negative) for 500k queries from Bing search engine	17,579,773
GOOAQ: Open Question Answering with Diverse Answer Types (query, answer) pairs for 3M Google queries and Google featured snippet	3,012,496
Amazon-QA (Question, Answer) pairs from Amazon product pages	2,448,839
Yahoo Answers (Title, Answer) pairs from Yahoo Answers	1,198,260
Yahoo Answers (Question, Answer) pairs from Yahoo Answers	681,164
Yahoo Answers (Title, Question) pairs from Yahoo Answers	659,896
SearchQA (Question, Answer) pairs for 140k questions, each with Top5 Google snippets on that question	582,261
ELI5 (Question, Answer) pairs from Reddit ELI5 (explainlikeimfive)	325,475
Stack Exchange Duplicate questions pairs (titles)	304,525
Quora Question Triplets (Question, Duplicate_Question, Hard_Negative) triplets for Quora Questions Pairs dataset	103,663
Natural Questions (NQ) (Question, Paragraph) pairs for 100k real Google queries with relevant Wikipedia paragraph	100,231
SQuAD2.0 (Question, Paragraph) pairs from SQuAD2.0 dataset	87,599
TriviaQA (Question, Evidence) pairs	73,346
Total	214,988,242