Papers

To ease the reproduction of papers, and the upload of full models on HuggingFace, the xpmir.papers package can be used.

Configuration

Papers experimental parameters are defined by data classes. The main one is xpmir.papers.helpers.PaperExperiment that defines an id (for experimaestro), a title and a description. These informations can be used e.g. when uploading the trained models on HuggingFace.

xpmir.papers.helpers.PaperExperiment: alias of ConfigurationBase

class xpmir.papers.helpers.NeuralIRExperiment(*, id: str = '???', file: str = 'experiment', parent: str | None = None, title: str = '', subtitle: str = '', paper: str = '', description: str = '', gpu: bool = True, seed: int = 0)[source]

Bases: ConfigurationBase

Settings most neural IR experiments

description: str: Description of the experiment

file: str: Relative path of the file containing a run function

gpu: bool: Use GPU for computation

id: str

ID of the experiment

This ID is used by experimaestro when running as the experiment.

paper: str: Source paper for this experiment

parent: str | None: Relative path of a YAML file that should be merged

seed: int: The seed used for experiments

subtitle: str: Allows to give some more details about the experiment

title: str: Short description of the experiment

Helpers

Pipelines factorize the code necessary to run some experiments: for instance, training re-rankers on MS-Marco is usually performed with similar training data, evaluation datasets.

Implemented papers

This page give some information about reproduction of papers based on XPMIR.

xpmir/cross-encoders:
- monoBERT