Top AI Repos — open-source AI, indexed and scored
Top AI Repos tracks AI repositories on GitHub and answers two different questions about each one: is it moving right now, and would you bet a product on it.
Top AI Repos tracks AI repositories on GitHub and answers two different questions about each one: is it moving right now, and would you bet a product on it.
Training & evaluation library for text-based neural re-ranking and dense retrieval models built with PyTorch
| Date | Stars |
|---|---|
| 2026-07-24 | 265 |
| 2026-07-25 | 265 |
| 2026-07-28 | 265 |
| 2026-07-30 | 265 |
| 2026-08-06 | 265 |
Today
— stars today
This week
— stars this week
This month
— stars this month
Momentum
0.0
growth rate 0.00%/day
# Welcome to Matchmaker 👋 Matchmaker is a research library for rapid training, evaluation, and analysis of text-based neural re-ranking and retrieval models built with PyTorch. Initially created to support Transformer-Kernel research, now evolved to a general library, with broad support for knowledge distillation. ## Features * 🔎 **Train a diverse set of neural IR models** including re-ranking & dense retrieval models based on HuggingFace transformers, TK-style models, and non-Bert neural baselines. * Wide range of configurations available * Static & dynamic sampling of batches (including TAS-Balanced) * Support for mixed precision (fp16) & multi-GPU training * 🧙♂️ **Knowledge distillation** with support for static + dynamic teachers; cross-architecture students; pairwise, in-batch negatives, dual-supervision * Including our Margin-MSE and many other loss options * 📄 **Evaluate models** with common IR metrics for multiple query sets for re-ranking and retrieval workflows * Efficiency metric tracking built in to everything * Including re-ranking depth evaluation * Highly optimized dense retrieval encoding, indexing, & search pipeline * 📚 **Experiment management** with folder based experiments, config & code tracking and the possibility to stack experiments one after another * Tensorboard files with dozen+ IR-specific telemetry statistics per experiment * Export models to the HuggingFace model hub If you have little experience with neural IR, we recommend you take a look at our (completely open & free) master level university course on *Advanced Information Retrieval* (Summer of 2021) at: https://github.com/sebastian-hofstaetter/teaching 🧙♂️ *Disclaimer: This library is created for research purposes, so it might break at any point. If you have problems, feel free to create an issue. If you are looking for a better tested, production ready search system - please have a look at [Pyserini](https://github.com/castorini/pyserini) or Vespa.ai* ## Getting started ### Requirements We recommend using a fresh conda environment with Python 3.8 *(can't use 3.9 atm, because of faiss)* ``` conda create -n matchmaker python=3.8 conda activate matchmaker ``` Then cd to the root folder of this repo, activate the conda environment, and install faiss & pytorch via conda. *We have to install faiss separately, because it does not have official pypi packages* ``` conda install --file conda-requirements.txt -c conda-forge -c pytorch ``` Then install the rest of the dependencies (allennlp, huggingface, ...) via pip install of the pip-requirements.txt ``` pip install -r pip-requirements.txt ``` **[Optional]** If you want to use the onnx runtime (currently implemented for dense_retrieval.py) you need to manually install cuDNN 8.0 (as of June 2021) which is a bit of a pain, but worth it :) Go to the NVIDIA website to download cuDNN 8.0 for your system (https://developer.nvidia.com/rdp/cudnn-archive), extract it and copy into the conda path (thanks to: https://stackoverflow.com/a/65335472): ``` cp cuda/include/cudnn*.h <conda environment path>/include cp cuda/lib64/libcudnn* <conda environment path>/lib ``` > For more info on onnx see: https://github.com/microsoft/onnxruntime/tree/master/onnxruntime/python/tools/transformers & https://medium.com/microsoftazure/accelerate-your-nlp-pipelines-using-hugging-face-transformers-and-onnx-runtime-2443578f4333 **Now you should be good to go!** Next up is matchmaker's data input format & depending on your interests we have some getting started guides below. ### Input Data Format In general every data input file in matchmaker is a .tsv file with tabs separating columns, as found in MSMARCO v1. We choose this format, because it allows us to easily open and operate on files, and changing tabs in the text does not alter the meaning of it. If you have a collection in another format such as older TREC collections or CORD-19 have a look in the *preprocessing/convert_formats* fold
Excerpt of 14,652 characters
Read on GitHubWould you bet a product on this? Bounded 0–100 and slow moving.
matched fp:45a8e9d1bec9ccf0, topic:knowledge-distillation, readme:knowledge distillation
matched fp:45a8e9d1bec9ccf0, topic:pytorch
matched fp:45a8e9d1bec9ccf0, topic:transformer