A comprehensive toolkit for evaluating NLP experiments offering various automated metrics. Jury offers a smooth and easy-to-use interface. It uses a more advanced version of evaluate design for underlying metric computation, so that adding custom metric is easy as extending proper class.

Main advantages that Jury offers are:

• Easy to use for any NLP project.
• Unified structure for computation input across all metrics.
• Calculate many metrics at once.
• Metrics calculations can be handled concurrently to save processing time.
• It seamlessly supports evaluation for multiple predictions/multiple references.

To see more, check the official Jury blog post.

• (2024.05.29) Retraction Watch Post regarding retraction of a paper has been posted. The plagiarised paper has been retracted.
• (2023.10.03) Jury paper is out currently is on arxiv. Please cite this paper if your work use Jury, and if your publication material will be submitted to the venues after this date.
• (2023.07.30) Public notice: You can reach our official Public Notice document that poses a claim about plagiarism of the work, jury, presented in this codebase.

The table below shows the current support status for available metrics.

* Placeholder for the rest of the metrics available in evaluate package apart from those which are present in the table.

Notes

• The entry ✔️ represents that full Jury support is available meaning that all combinations of input types (single prediction & single reference, single prediction & multiple references, multiple predictions & multiple references) are supported

• The entry ✅ means that this metric is supported (for Jury through the evaluate), so that it can (and should) be used just like the evaluate metric as instructed in evaluate implementation although unfortunately full Jury support for those metrics are not yet available.

For the request of a new metric please open an issue providing the minimum information. Also, PRs addressing new metric supports are welcomed :).

Through pip,

pip install jury

or build from source,

git clone https://github.com/obss/jury.git
cd jury
python setup.py install

NOTE: There may be malfunctions of some metrics depending on sacrebleu package on Windows machines which is mainly due to the package pywin32. For this, we fixed pywin32 version on our setup config for Windows platforms. However, if pywin32 causes trouble in your environment we strongly recommend using conda manager install the package as conda install pywin32.

It is only two lines of code to evaluate generated outputs.

from jury import Jury

scorer = Jury()
predictions = [
    ["the cat is on the mat", "There is cat playing on the mat"], 
    ["Look!    a wonderful day."]
]
references = [
    ["the cat is playing on the mat.", "The cat plays on the mat."], 
    ["Today is a wonderful day", "The weather outside is wonderful."]
]
scores = scorer(predictions=predictions, references=references)

Specify metrics you want to use on instantiation.

scorer = Jury(metrics=["bleu", "meteor"])
scores = scorer(predictions, references)

You can directly import metrics from jury.metrics as classes, and then instantiate and use as desired.

from jury.metrics import Bleu

bleu = Bleu.construct()
score = bleu.compute(predictions=predictions, references=references)

The additional parameters can either be specified on compute()

from jury.metrics import Bleu

bleu = Bleu.construct()
score = bleu.compute(predictions=predictions, references=references, max_order=4)

, or alternatively on instantiation

from jury.metrics import Bleu
bleu = Bleu.construct(compute_kwargs={"max_order": 1})
score = bleu.compute(predictions=predictions, references=references)

Note that you can seemlessly access both jury and evaluate metrics through jury.load_metric.

import jury

bleu = jury.load_metric("bleu")
bleu_1 = jury.load_metric("bleu", resulting_name="bleu_1", compute_kwargs={"max_order": 1})
# metrics not available in `jury` but in `evaluate`
wer = jury.load_metric("competition_math") # It falls back to `evaluate` package with a warning

You can specify predictions file and references file paths and get the resulting scores. Each line should be paired in both files. You can optionally provide reduce function and an export path for results to be written.

jury eval --predictions /path/to/predictions.txt --references /path/to/references.txt --reduce_fn max --export /path/to/export.txt

You can also provide prediction folders and reference folders to evaluate multiple experiments. In this set up, however, it is required that the prediction and references files you need to evaluate as a pair have the same file name. These common names are paired together for prediction and reference.

jury eval --predictions /path/to/predictions_folder --references /path/to/references_folder --reduce_fn max --export /path/to/export.txt

If you want to specify metrics, and do not want to use default, specify it in config file (json) in metrics key.

{
  "predictions": "/path/to/predictions.txt",
  "references": "/path/to/references.txt",
  "reduce_fn": "max",
  "metrics": [
    "bleu",
    "meteor"
  ]
}

Then, you can call jury eval with config argument.

jury eval --config path/to/config.json

You can use custom metrics with inheriting jury.metrics.Metric, you can see current metrics implemented on Jury from jury/metrics. Jury falls back to evaluate implementation of metrics for the ones that are currently not supported by Jury, you can see the metrics available for evaluate on evaluate/metrics.

Jury itself uses evaluate.Metric as a base class to drive its own base class as jury.metrics.Metric. The interface is similar; however, Jury makes the metrics to take a unified input type by handling the inputs for each metrics, and allows supporting several input types as;

• single prediction & single reference
• single prediction & multiple reference
• multiple prediction & multiple reference

As a custom metric both base classes can be used; however, we strongly recommend using jury.metrics.Metric as it has several advantages such as supporting computations for the input types above or unifying the type of the input.

from jury.metrics import MetricForTask

class CustomMetric(MetricForTask):
    def _compute_single_pred_single_ref(
        self, predictions, references, reduce_fn = None, **kwargs
    ):
        raise NotImplementedError

    def _compute_single_pred_multi_ref(
        self, predictions, references, reduce_fn = None, **kwargs
    ):
        raise NotImplementedError

    def _compute_multi_pred_multi_ref(
            self, predictions, references, reduce_fn = None, **kwargs
    ):
        raise NotImplementedError

For more details, have a look at base metric implementation jury.metrics.Metric

PRs are welcomed as always :)

git clone https://github.com/obss/jury.git
cd jury
pip install -e ".[dev]"

Also, you need to install the packages which are available through a git source separately with the following command. For the folks who are curious about "why?"; a short explaination is that PYPI does not allow indexing a package which are directly dependent on non-pypi packages due to security reasons. The file requirements-dev.txt includes packages which are currently only available through a git source, or they are PYPI packages with no recent release or incompatible with Jury, so that they are added as git sources or pointing to specific commits.

pip install -r requirements-dev.txt

To tests simply run.

python tests/run_tests.py

To check code style,

python tests/run_code_style.py check

To format codebase,

python tests/run_code_style.py format

If you use this package in your work, please cite it as:

@misc{cavusoglu2023jury,
  title={Jury: A Comprehensive Evaluation Toolkit}, 
  author={Devrim Cavusoglu and Ulas Sert and Secil Sen and Sinan Altinuc},
  year={2023},
  eprint={2310.02040},
  archivePrefix={arXiv},
  primaryClass={cs.CL},
  doi={10.48550/arXiv.2310.02040}
}

We use the GitHub Issue Tracker to track issues in general. Issues can be bug reports, feature requests or implementation of a new metric type. Please refer to the related issue template for opening new issues.

Licensed under the MIT License.