MVDR_pruning

KO | EN

gitlite — search

#python #hacktoberfest #typescript #llm #golang #rust #deep-learning #linux #macos #cli #kubernetes #security #ios #mcp

(★ 20)

A project which does the ColBERT pruning based on the LP or L1 norm

MVDR_pruning 최신버젼 다운로드

최종 버전 다운로드 (.zip)

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

Towards Lossless Token Pruning in Late-Interaction Retrieval Models
===

This repository contains all the details, configurations and scripts necessary in order to reproduce the experiments led in the context of the paper *Towards Lossless Token Pruning in Late-Interaction Retrieval Models* by Yuxuan Zong and Benjamin Piwowarski, that was accepted at SIGIR 2025.

## Table of Contents

* [Installation](#installation)
* [Examples](#examples)
* [Experiment Details](#experiment-details)
* [Contact](#contact)
* [Citation](#citation)

## Installation

The repository is build based on the [experimaestro-ir](https://github.com/experimaestro/experimaestro-ir) framework, which is a repository providing many basic key components in doing IR experiments, including learning, indexing, evaluating, etc. Our dataset access is done under the [ir-datasets](https://github.com/allenai/ir_datasets/) package, through the interface of [datamaestro](https://github.com/experimaestro/datamaestro_text).

The repository can be installed locally by cloning it before running `pip install -r requirements.txt` (Note that in order to get everything to work you might need to ponctually install other libs).

When setting the experimaestro experiments, you also need to prepare a `launcher.py` under the folder `~/.config/experimaestro`. Examples could be found [here](https://experimaestro-python.readthedocs.io/en/latest/launchers/#launcher-file-since-124)

## Experiment Details

For the hyperparameters of the model, please find the details in the `.yaml` files. We show several important ones in the following part:

### Regularization Hyperparameters
The regularization hyperparameters are defined under `RegularizationCoeffConfiguration` class of the `experiments/configuration.py` file. The one used for the experiments are defined under the corresponding hierarchy in the `.yaml` file.

### Pruning Hyperparameters
The pruning hyperparameters are defined under `PruningConfig` class of the `experiments/configuration.py` file. The one used for the experiments are defined under the corresponding hierarchy in the `.yaml` file. We use lists to represent the possible pruning threshold in order to gather several different experiments in the same experimental file.

### Dataset

#### Distillation Samples

Please replace the `distil_data_path` in the `.yaml` to the path where where you store the distillation samples provided by the [ColBERTv2](https://github.com/stanford-futuredata/ColBERT). You can download it by:

```unix
wget https://huggingface.co/colbert-ir/colbertv2.0_msmarco_64way/resolve/main/examples.json?download=true
```

#### Evaluation
We make use of the package [`ir-datasets`](https://ir-datasets.com/) together with [`datamaestro`](https://github.com/experimaestro/datamaestro_text) to facilitate the access of the datasets in our experiments. In `.yaml` file, we only an example of using the *TREC-COVID* and *LoTTE-recreation* for the out-of-domain evaluation and commented out the other dataset used in our experiments.

To use other dataset, please use the following command to check the availability:
```unix
datamaestro search <dataset_name>
```

## Experiments Reproduction

In this part, we describe how to reproduce the In-domain and Out-of-domain Evaluation Results presented in the paper. Following the [experimaestro framework](https://experimaestro-python.readthedocs.io/en/latest/), the whole pipeline of indexing, training, evaluating are defined under *Experiment* scripts, controlled by configuration file written in YAML. Examples of such files are available under the [experiments](experiments) folder of the repository. We can run the following command to launch the experiments:

````unix
experimaestro run-experiment experiments/normal.yaml --workdir /your/working/directory/
````

Alternatively, if you want to make sure that everything is setup correctly before starting your experiment, you can add the option `--run-mode dry-run` to the command above (Go over the experimental plan without launching the task).

The `experiments/normal.yaml` already contains the hyperparameters we need to reproduce the in-domain evaluations. To evaluate on out-of-domain dataset, please set the `evaluate_in_domain` to `false` in the `.yaml` file.

## Contact

Please feel free to email Yuxuan or the academic' supervisor Benjamin at (name).(surname)@isir.upmc.fr

## Citation

```
@article{zong2025towards,
  title={Towards Lossless Token Pruning in Late-Interaction Retrieval Models},
  author={Zong, Yuxuan and Piwowarski, Benjamin},
  journal={arXiv preprint arXiv:2504.12778},
  year={2025}
}
```

MVDR_pruning

MVDR_pruning 최신버젼 다운로드

오늘의 추천 프로젝트

Was this content helpful?

Recent Feedback

MVDR_pruning

MVDR_pruning 최신버젼 다운로드

오늘의 추천 프로젝트

Was this content helpful?

Recent Feedback

Download README