Team Ai
Modelpublic

ChristianAzinn/labse-gguf

sourceHugging Faceapache-2.0updated 2y agoView on Hugging Face
5likes546downloads
Model Card

labse-gguf

Model creator: sentence-transformers

Original model: labse

Original Description

The language-agnostic BERT sentence embedding encodes text into high dimensional vectors. The model is trained and optimized to produce similar representations exclusively for bilingual sentence pairs that are translations of each other. So it can be used for mining for translations of a sentence in a larger corpus.

Description

This repo contains GGUF format files for the labse embedding model.

These files were converted and quantized with llama.cpp PR 5500, commit 34aa045de, on a consumer RTX 4090.

This model supports up to 512 tokens of context.

Compatibility

These files are compatible with llama.cpp as of commit 4524290e8, as well as LM Studio as of version 0.2.19.

Meta-information

Explanation of quantisation methods

<details> <summary>Click to see details</summary> The methods available are:

  • —GGMLTYPEQ2_K - "type-1" 2-bit quantization in super-blocks containing 16 blocks, each block having 16 weight. Block scales and mins are quantized with 4 bits. This ends up effectively using 2.5625 bits per weight (bpw)
  • —GGMLTYPEQ3_K - "type-0" 3-bit quantization in super-blocks containing 16 blocks, each block having 16 weights. Scales are quantized with 6 bits. This end up using 3.4375 bpw.
  • —GGMLTYPEQ4_K - "type-1" 4-bit quantization in super-blocks containing 8 blocks, each block having 32 weights. Scales and mins are quantized with 6 bits. This ends up using 4.5 bpw.
  • —GGMLTYPEQ5K - "type-1" 5-bit quantization. Same super-block structure as GGMLTYPEQ4K resulting in 5.5 bpw
  • —GGMLTYPEQ6_K - "type-0" 6-bit quantization. Super-blocks with 16 blocks, each block having 16 weights. Scales are quantized with 8 bits. This ends up using 6.5625 bpw Refer to the Provided Files table below to see what files use which methods, and how. </details>

Provided Files

NameQuant methodBitsSizeMax RAM requiredUse case
NameQuant methodBitsSizeUse case
labse.Q2_K.ggufQ2_K2364 MBsmallest, significant quality loss - not recommended for most purposes
labse.Q3_K_S.ggufQ3KS3368 MBvery small, high quality loss
labse.Q3_K_M.ggufQ3KM3374 MBvery small, high quality loss
labse.Q3_K_L.ggufQ3KL3379 MBsmall, substantial quality loss
labse.Q4_0.ggufQ4_04379 MBlegacy; small, very high quality loss - prefer using Q3KM
labse.Q4_K_S.ggufQ4KS4380 MBsmall, greater quality loss
labse.Q4_K_M.ggufQ4KM4384 MBmedium, balanced quality - recommended
labse.Q5_0.ggufQ5_05390 MBlegacy; medium, balanced quality - prefer using Q4KM
labse.Q5_K_S.ggufQ5KS5390 MBlarge, low quality loss - recommended
labse.Q5_K_M.ggufQ5KM5392 MBlarge, very low quality loss - recommended
labse.Q6_K.ggufQ6_K6401 MBvery large, extremely low quality loss
labse.Q8_0.ggufQ8_08515 MBvery large, extremely low quality loss - recommended
labse.Q8_0.gguffp1616955 MBenormous, pretty much the original model - not recommended
labse.Q8_0.gguffp32321.89 GBenormous, pretty much the original model - not recommended

Examples

Example Usage with llama.cpp

To compute a single embedding, build llama.cpp and run:

shell
./embedding -ngl 99 -m [filepath-to-gguf].gguf -p 'search_query: What is TSNE?'

You can also submit a batch of texts to embed, as long as the total number of tokens does not exceed the context length. Only the first three embeddings are shown by the embedding example.

texts.txt:

search_query: What is TSNE?
search_query: Who is Laurens Van der Maaten?

Compute multiple embeddings:

shell
./embedding -ngl 99 -m [filepath-to-gguf].gguf -f texts.txt

Example Usage with LM Studio

Download the 0.2.19 beta build from here: Windows MacOS Linux

Once installed, open the app. The home should look like this:

image/png

Search for either "ChristianAzinn" in the main search bar or go to the "Search" tab on the left menu and search the name there.

image/png

Select your model from those that appear (this example uses bge-small-en-v1.5-gguf) and select which quantization you want to download. Since this model is pretty small, I recommend Q8_0, if not f16/32. Generally, the lower you go in the list (or the bigger the number gets), the larger the file and the better the performance.

image/png

You will see a green checkmark and the word "Downloaded" once the model has successfully downloaded, which can take some time depending on your network speeds.

image/png

Once this model is finished downloading, navigate to the "Local Server" tab on the left menu and open the loader for text embedding models. This loader does not appear before version 0.2.19, so ensure you downloaded the correct version.

image/png

Select the model you just downloaded from the dropdown that appears to load it. You may need to play with configuratios in the right-side menu, such as GPU offload if it doesn't fit entirely into VRAM.

image/png

All that's left to do is to hit the "Start Server" button:

image/png

And if you see text like that shown below in the console, you're good to go! You can use this as a drop-in replacement for the OpenAI embeddings API in any application that requires it, or you can query the endpoint directly to test it out.

image/png

Example curl request to the API endpoint:

shell
curl http://localhost:1234/v1/embeddings \
  -H "Content-Type: application/json" \
  -d '{
    "input": "Your text string goes here",
    "model": "model-identifier-here"
  }'

For more information, see the LM Studio text embedding documentation.

Acknowledgements

Thanks to the LM Studio team and everyone else working on open-source AI.

This README is inspired by that of nomic-ai-embed-text-v1.5-gguf, another excellent embedding model, and those of the legendary TheBloke.