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

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:

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

all-MiniLM-L12-v2

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.

---

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.