Metadata-Version: 2.1
Name: distfuse
Version: 0.1.4
Summary: Compute DistFuse similarity scores from embedding models and APIs
Home-page: https://github.com/gentaiscool/distfuse
Author: Genta Indra Winata
Author-email: gentaindrawinata@gmail.com
License: Apache 2.0 License
Project-URL: Bug Tracker, https://github.com/gentaiscool/distfuse/issues
Classifier: Programming Language :: Python :: 3
Classifier: License :: OSI Approved :: MIT License
Classifier: Operating System :: OS Independent
Requires-Python: >=3.7
Description-Content-Type: text/markdown
License-File: LICENSE
Requires-Dist: scikit-learn >=1.5.0
Requires-Dist: sentence-transformers
Requires-Dist: numpy
Requires-Dist: openai
Requires-Dist: cohere

# DistFuse

DistFuse is a library to calculate similarity scores between two collections of text sequences encoded using transformer models. This library allows combining multiple models, including Hugging Face encoder models and embed APIs from Cohere and OpenAI. 

## Use Cases
This is the same implementation of DistFuse from the [MINERS paper](https://arxiv.org/pdf/2406.07424). This library is useful for bitext mining, dense retrieval, retrieval-based classification, and retrieval-augmented generation (RAG).

## Table of Contents

- [Install](#install)
- [Usage](#usage)
- [Reference](#reference)
- [How to Contribute?](#-how-to-contribute)

## Install from pypi (stable)
```
pip install distfuse
```

## Install from source (latest)
```
git clone --recursive https://github.com/gentaiscool/distfuse
pip install .
```

## Usage
We support `hf` (Hugging Face models), and APIs, such as `cohere`, and `openai`. For `dist_measure`, we support `cosine`, `euclidean`, and `manhattan`. If you are planning to use API models, please pass the appropriate token to `openai_token` or `cohere_token`. To use more than one model, add the model information to `model_checkpoints` and the weight to `weights`. There is no limit to the number of models you can use.

### Generate Pairwise Scores
If you want to generate pairwise scores between two lists, you can call `score_pairs`. Here are the examples:

e.g., DistFuse with 2 models.
```python
from distfuse import DistFuse

model_checkpoints = [["sentence-transformers/LaBSE", "hf"], ["sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2", "hf"]]
weights = [1, 1]
dist_measure = "cosine" # cosine, euclidean, manhattan
model = DistFuse(model_checkpoints, weights, dist_measure=dist_measure, openai_token="", cohere_token="", device="cuda:0")

scores = model.score_pairs(["I like apple", "I like cats"], ["I like orange", "I like dogs"])
print(scores)
```

e.g., DistFuse with 3 models. 
```python
from distfuse import DistFuse

model_checkpoints = [["sentence-transformers/LaBSE", "hf"], ["sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2", "hf"], ["text-embedding-3-large", "openai"]]
weights = [1, 1, 1]
dist_measure = "cosine"
model = DistFuse(model_checkpoints, weights, dist_measure=dist_measure, openai_token="", cohere_token="", device="cuda:0")

scores = model.score_pairs(["I like apple", "I like cats"], ["I like orange", "I like dogs"])
print(scores)
```

e.g., DistFuse with 2 models and custom instruction.
```python
from distfuse import DistFuse

model_checkpoints = [["sentence-transformers/LaBSE", "hf"], ["intfloat/multilingual-e5-large-instruct", "hf"]]
weights = [1, 1]
dist_measure = "cosine" # cosine, euclidean, manhattan
instructions = ["", "Given a web search query, retrieve relevant passages that answer the query"]
model = DistFuse(model_checkpoints, weights, instructions, dist_measure=dist_measure, openai_token="", cohere_token="", device="cuda:0")

scores = model.score_pairs(["I like apple", "I like cats"], ["I like orange", "I like dogs"])
print(scores)
```

### Generate Predictions to Multi-reference Scores
If you want to generate scores to compare the distance between the predictions and multi-reference, you can call `score_reference` and pass a list of string as `predictions` with a size of `m` and a list of list of string as `references` with as size of `m` and `r`, where `r` is the length of the references. You can have a variable number of `r` for each sample. The lengths of first dimension the `predictions` and `references` have to be the same. Here are the examples:

e.g., DistFuse with 2 models. 
```python
from distfuse import DistFuse

model_checkpoints = [["sentence-transformers/LaBSE", "hf"], ["sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2", "hf"]]
weights = [1, 1]
dist_measure = "cosine"
model = DistFuse(model_checkpoints, weights, dist_measure=dist_measure, openai_token="", cohere_token="", device="cuda:0")

scores = model.score_references(predictions=["I like apple", "I like cats"], references=[["I like orange", "I like dogs"],["I like orange", "I like dogs"]])
print(scores)
```

## Reference
If you use any source codes included in this toolkit in your work, please cite the following papers [[1]](https://arxiv.org/pdf/2406.07424) [[2]](https://aclanthology.org/2023.ijcnlp-short.11.pdf).
```
@article{winata2024miners,
  title={MINERS: Multilingual Language Models as Semantic Retrievers},
  author={Winata, Genta Indra and Zhang, Ruochen and Adelani, David Ifeoluwa},
  journal={arXiv preprint arXiv:2406.07424},
  year={2024}
}
@inproceedings{winata2023efficient,
  title={Efficient Zero-Shot Cross-lingual Inference via Retrieval},
  author={Winata, Genta and Xie, Lingjue and Radhakrishnan, Karthik and Gao, Yifan and Preo{\c{t}}iuc-Pietro, Daniel},
  booktitle={Proceedings of the 13th International Joint Conference on Natural Language Processing and the 3rd Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics (Volume 2: Short Papers)},
  pages={93--104},
  year={2023}
}
```

## 🚀 How to Contribute?
Feel free to create [an issue](https://github.com/gentaiscool/distfuse/issues/) if you have any questions. And, create [a PR](https://github.com/gentaiscool/distfuse/pulls) for fixing bugs or adding improvements. 

If you are interested to create an extension of this work, feel free to reach out to [us](mailto:gentaindrawinata@gmail.com)!

Support our open source effort ⭐
