metadata

license: mit
task_categories:
  - text-classification
language:
  - en
pretty_name: sst2_cognitive-bias
size_categories:
  - 100K<n<1M
source_datasets:
  - sst2
dataset_info:
  features:
    - name: idx
      dtype: string
    - name: sentence
      dtype: string
    - name: label
      dtype: int64
    - name: dist
      dtype: string
    - name: shot1_idx
      dtype: string
    - name: shot1_sent
      dtype: string
    - name: shot1_label
      dtype: int64
    - name: shot2_idx
      dtype: string
    - name: shot2_sent
      dtype: string
    - name: shot2_label
      dtype: int64
    - name: shot3_idx
      dtype: string
    - name: shot3_sent
      dtype: string
    - name: shot3_label
      dtype: int64
    - name: shot4_idx
      dtype: string
    - name: shot4_sent
      dtype: string
    - name: shot4_label
      dtype: int64
    - name: few_shot_string
      dtype: string
    - name: few_shot_hard_string
      dtype: string
    - name: id
      dtype: int64
  splits:
    - name: train
      num_bytes: 286790625
      num_examples: 250000
  download_size: 47727501
  dataset_size: 286790625
splits:
  - name: train
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*

Dataset Card for cobie_sst2

This dataset is a modification of the original SST-2 dataset for LLM cognitive bias evaluation.

Language(s)

English (en)

Dataset Summary

The Stanford Sentiment Treebank is a corpus with fully labeled parse trees that allows for a complete analysis of the compositional effects of sentiment in language. The corpus is based on the dataset introduced by Pang and Lee (2005) and consists of 11,855 single sentences extracted from movie reviews. It was parsed with the Stanford parser and includes a total of 215,154 unique phrases from those parse trees, each annotated by 3 human judges.

Dataset Structure

The modifications carried out in the dataset are thought to evaluate cognitive biases in a few-shot setting and with two different task complexities. We make use of 25,000 instances from the original dataset, while the remaining ones serve as few-shot examples. Each instance is prompted with all possible unbalanced 4-shot distributions. To increase the original task complexity, we also introduce an additional neutral example between the first and last two examples.

Dataset Fields

idx: original sentence id, in the format <original_partition>_<original_id>.
sentence: test sentence.
label: sentiment of the test sentence, either "negative" (0) or "positive" (1).
dist: few-shot distribution (0000, 1111, 0001, 0010, 0100, 1000, 1110, 1101, 1011, 0111).
shot<n>_idx: original id of the example sentence, in the format <original_partition>_<original_id>.
shot<n>_sent: example sentence.
shot<n>_label: sentiment of the example sentence.
few_shot_string: string with all 4 shots the sentence is prompted with.
few_shot_hard_string: string with the same 4 shots and an additional neutral example between the first and last two to increase task complexity.

Supported Tasks and Leaderboards

sentiment-classification

Additional Information

Dataset Curators

Language Technologies Unit (LangTech) at the Barcelona Supercomputing Center.

This work has been promoted and financed by the Generalitat de Catalunya through the Aina project. This work is also funded by the Ministerio para la Transformación Digital y de la Función Pública and Plan de Recuperación, Transformación y Resiliencia - Funded by EU – NextGenerationEU within the framework of the project Desarrollo Modelos ALIA.

Licensing Information

This work is licensed under a MIT License (same as original).

Citation Information

@inproceedings{cobie,
  title={Cognitive Biases, Task Complexity, and Result Intepretability in Large Language Models},
  author={Mario Mina and Valle Ruiz-Fernández and Júlia Falcão and Luis Vasquez-Reina and Aitor Gonzalez-Agirre},
  booktitle={Proceedings of The 31st International Conference on Computational Linguistics (COLING)},
  year={2025 (to appear)}
}