LucasWeber/icl_consistency_test

The ICL consistency test

This 🤗 dataset provides data for the GenBench CBT task 'The ICL consistency test'. The ICL consistency test measures the consistency of LLM predictions on the same data points across many different equivalent prompting setups. The score in the associated metric (Cohen's kappa) can be understood as a measure of a model's prediction consistency in the face of task-irrelevant information.

For an easy evaluation of any 🤗 models, we refer to the code provided in the GenBench task. For in-depth information on the task, we refer to the associated publications (Weber et al., 2023,2023) and the respective GenBench doc.md.

Evaluation on the relevant metrics can be done via the example_evaluation.py script in the GenBench repository.

Dataset Description

Abstract: The ICL consistency test measures the consistency of LLM predictions on the same data points across many different prompting setups. Different setups are defined by "factors". On the one hand, factors can be specific attributes of the used prompt (e.g. the number of examples the model is presented with ["n_shots"] or the type of instructions that were used to wrap a specific datapoint ["Instructions"]). On the other hand, the analysis can also be augmented by factors that are related to the way a model is evaluated (e.g. whether a model is calibrated) or the type of model that is evaluated (e.g. the number of parameters or instructions tuning). These external factors can be added to the analysis by using the task.add_factor() method. The output metric is Cohen's kappa for each factor across all different conditions. A kappa value close to 1 indicates that the factors do not change the model prediction, while a factor close to 0 strongly changes model predictions. The ICL consistency test has two subtasks, one evaluating the ANLI-dataset (Nie et al., 2019); the other the MNLI-dataset (Wang et al., 2017).

Size: Each subtask contains 57600 when using the full 600 data_IDs. The user can choose to reduce the number of evaluated data_IDs.

• Curated by:
  • resampling and arrangement was done by Weber et al., 2023,2023;
  • original data were curated by Nie et al., 2019 (ANLI) and Wang et al., 2017 (MNLI);
  • templates were curated by Bach et al., 2022 (promptsource).
• Language: English

Dataset Sources (basic links)

• Repository: Data files on github.
• Paper: Weber et al., 2023,2023.
• Demo: Find pre-implemented code to evaluate any 🤗 model on github.

Uses

In prompting, models are sensitive to task-irrelevant information in their prompt. This test can be used to quantify this sensitivity of any 🤗 model. The ICL consistency test does this by measuring a model's prediction consistency across many different semantically equivalent prompting setups.

Dataset Structure

[TBA]

Dataset Creation

The data is a sample from the MNLI and ANLI datasets as well as prompt templates from promptsource. Please refer to the original publications's documentation for detailed information on dataset creation.

Bias, Risks, and Limitations

This dataset contains data from the MNLI and ANLI datasets and adheres to the same biases, risks and limitations.

Recommendations

We identify the following limitations of the consistency test:

1. The number of factors is limited and does not cover all possible factors that might influence the predictions. We limited ourselves to factors we deem relevant, to ensure fast evaluation.

2. Currently, the test is only implemented for the ANLI- and MNLI-datasets.

3. Factors that are external to the dataset but should be considered in the analysis (e.g. instruction tuning or calibration) have to be manually added by the user using the task.add_factor() method (please use the GenBench implementation of the dataset. You can find it on github).

Citation

This dataset was used in the following publications. If you use it, please consider citing the following references:

BibTeX:

@inproceedings{weber2023mind,
  title={Mind the instructions: a holistic evaluation of consistency and interactions in prompt-based learning},
  author={Weber, Lucas and Bruni, Elia and Hupkes, Dieuwke},
  booktitle={Proceedings of the 27th Conference on Computational Natural Language Learning (CoNLL)},
  pages={294--313},
  year={2023}
}

@article{weber2023icl,
  title={The ICL Consistency Test},
  author={Weber, Lucas and Bruni, Elia and Hupkes, Dieuwke},
  journal={arXiv preprint arXiv:2312.04945},
  year={2023}
}

Dataset Card Authors

Lucas Weber

Dataset Card Contact

lucasweber000@gmail.com