bias-amplified-splits/qqp

---
license: cc-by-4.0
dataset_info:
- config_name: minority_examples
features:
- name: question1
dtype: string
- name: question2
dtype: string
- name: label
dtype:
class_label:
names:
'0': not_duplicate
'1': duplicate
- name: idx
dtype: int32
splits:
- name: train.biased
num_bytes: 42391456
num_examples: 297735
- name: train.anti_biased
num_bytes: 8509364
num_examples: 66111
- name: validation.biased
num_bytes: 4698206
num_examples: 32968
- name: validation.anti_biased
num_bytes: 955548
num_examples: 7462
download_size: 70726976
dataset_size: 56554574
- config_name: partial_input
features:
- name: question1
dtype: string
- name: question2
dtype: string
- name: label
dtype:
class_label:
names:
'0': not_duplicate
'1': duplicate
- name: idx
dtype: int32
splits:
- name: train.biased
num_bytes: 42788212
num_examples: 297735
- name: train.anti_biased
num_bytes: 8112608
num_examples: 66111
- name: validation.biased
num_bytes: 4712327
num_examples: 33084
- name: validation.anti_biased
num_bytes: 941427
num_examples: 7346
download_size: 70726976
dataset_size: 56554574
task_categories:
- text-classification
language:
- en
pretty_name: Quora Questions Pairs
---

# Dataset Card for Bias-amplified Splits for QQP

## Table of Contents
- [Table of Contents](#table-of-contents)
- [Dataset Description](#dataset-description)
- [Dataset Summary](#dataset-summary)
- [Dataset Structure](#dataset-structure)
- [Data Instances](#data-instances)
- [Data Fields](#data-fields)
- [Data Splits](#data-splits)
- [Dataset Creation](#dataset-creation)
- [Curation Rationale](#curation-rationale)
- [Annotations](#annotations)
- [Considerations for Using the Data](#considerations-for-using-the-data)
- [Social Impact of Dataset](#social-impact-of-dataset)
- [Discussion of Biases](#discussion-of-biases)
- [Additional Information](#additional-information)
- [Dataset Curators](#dataset-curators)
- [Citation Information](#citation-information)

## Dataset Description

- **Repository:** [Fighting Bias with Bias repo](https://github.com/schwartz-lab-nlp/fight-bias-with-bias)
- **Paper:** [arXiv](https://arxiv.org/abs/2305.18917)
- **Point of Contact:** [Yuval Reif](mailto:[e-mail])
- **Original Dataset's Paper:** [GLUE](https://arxiv.org/abs/1804.07461)

### Dataset Summary

Bias-amplified splits is a novel evaluation framework to assess model robustness, by amplifying dataset biases in the training data and challenging models to generalize beyond them. This framework is defined by a bias-amplified training set and a hard, anti-biased test set, which we automatically extract from existing datasets using model-based methods.

Our experiments show that the identified anti-biased examples are naturally challenging for models, and moreover, models trained on bias-amplified data exhibit dramatic performance drops on anti-biased examples, which are not mitigated by common approaches to improve generalization.

Here we apply our framework to the Quora Question Pairs dataset (QQP), a dataset composed of question pairs where the task is to determine if the questions are paraphrases of each other (have the same meaning).

Our evaluation framework can be applied to any existing dataset, even those considered obsolete, to test model robustness. We hope our work will guide the development of robust models that do not rely on superficial biases and correlations.

#### Evaluation Results (DeBERTa-large)

##### For splits based on minority examples:

| Training Data \ Test Data | Original test | Anti-biased test |
|---------------------------|---------------|------------------|
| Original training split | 93.0 | 77.6 |
| Biased training split | 87.0 | 36.8 |

##### For splits based on partial-input model:

| Training Data \ Test Data | Original test | Anti-biased test |
|---------------------------|---------------|------------------|
| Original training split | 93.0 | 81.3 |
| Biased training split | 90.3 | 63.9 |

#### Loading the Data

```
from datasets import load_dataset

# choose which bias detection method to use for the bias-amplified splits: either "minority_examples" or "partial_input"
dataset = load_dataset("bias-amplified-splits/qqp", "minority_examples")

# use the biased training split and anti-biased test split
train_dataset = dataset['train.biased']
eval_dataset = dataset['validation.anti_biased']
```

## Dataset Structure

### Data Instances

Data instances are taken directly from QQP (GLUE version), and re-split into biased and anti-biased subsets. Here is an example of an instance from the dataset:
```
{
"idx": 56,
"question1": "How do I buy used car in India?",
"question2": "Which used car should I buy in India?",
"label": 0
}
```

### Data Fields

- `idx`: unique identifier for the example within its original data splits (e.g., validation set)
- `question1`: a question asked on Quora
- `question2`: a question asked on Quora
- `label`: one of `0` and `1` (`not duplicate` and `duplicate`)

### Data Splits

Bias-amplified splits require a method to detect *biased* and *anti-biased* examples in datasets. We release bias-amplified splits based created with each of these two methods:
- **Minority examples**: A novel method we introduce that leverages representation learning and clustering for identifying anti-biased *minority examples* (Tu et al., 2020)—examples that defy common statistical patterns found in the rest of the dataset.
- **Partial-input baselines**: A common method for identifying biased examples containing annotation artifacts in a dataset, which examines the performance of models that are restricted to using only part of the input. Such models, if successful, are bound to rely on unintended or spurious patterns in the dataset.

Using each of the two methods, we split each of the original train and test splits into biased and anti-biased subsets. See the [paper](https://arxiv.org/abs/2305.18917) for more details.

#### Minority Examples

| Dataset Split | Number of Instances in Split |
|--------------------------|------------------------------|
| Train - biased | 297735 |
| Train - anti-biased | 66111 |
| Validation - biased | 32968 |
| Validation - anti-biased | 7462 |

#### Partial-input Baselines

| Dataset Split | Number of Instances in Split |
|--------------------------|------------------------------|
| Train - biased | 297735 |
| Train - anti-biased | 66111 |
| Validation - biased | 33084 |
| Validation - anti-biased | 7346 |

## Dataset Creation

### Curation Rationale

NLP models often rely on superficial cues known as *dataset biases* to achieve impressive performance, and can fail on examples where these biases do not hold. To develop more robust, unbiased models, recent work aims to filter bisased examples from training sets. We argue that in order to encourage the development of robust models, we should in fact **amplify** biases in the training sets, while adopting the challenge set approach and making test sets anti-biased. To implement our approach, we introduce a simple framework that can be applied automatically to any existing dataset to use it for testing model robustness.

### Annotations

#### Annotation process

No new annotations are required to create bias-amplified splits. Existing data instances are split into *biased* and *anti-biased* splits based on automatic model-based methods to detect such examples.

## Considerations for Using the Data

### Social Impact of Dataset

Bias-amplified splits were created to promote the development of robust NLP models that do not rely on superficial biases and correlations, and provide more challenging evaluation of existing systems.

### Discussion of Biases

We propose to use bias-amplified splits to complement benchmarks with challenging evaluation settings that test model robustness, in addition to the dataset’s main training and test sets. As such, while existing dataset biases are *amplified* during training with bias-amplified splits, these splits are intended primarily for model evaluation, to expose the bias-exploiting behaviors of models and to identify more robsut models and effective robustness interventions.

## Additional Information

### Dataset Curators

Bias-amplified splits were introduced by Yuval Reif and Roy Schwartz from the [Hebrew University of Jerusalem](https://schwartz-lab-huji.github.io).

QQP data was released by Quora and released under the GLUE benchmark.

### Citation Information

```
@misc{reif2023fighting,
title = "Fighting Bias with Bias: Promoting Model Robustness by Amplifying Dataset Biases",
author = "Yuval Reif and Roy Schwartz",
month = may,
year = "2023",
url = "https://arxiv.org/pdf/2305.18917",
}
```

Source dataset:
```
@inproceedings{wang2019glue,
title={{GLUE}: A Multi-Task Benchmark and Analysis Platform for Natural Language Understanding},
author={Wang, Alex and Singh, Amanpreet and Michael, Julian and Hill, Felix and Levy, Omer and Bowman, Samuel R.},
note={In the Proceedings of ICLR.},
year={2019}
}
```

原始信息汇总

数据集卡片 for Bias-amplified Splits for QQP

数据集描述

数据集摘要

Bias-amplified splits 是一种新颖的评估框架，用于通过在训练数据中放大数据集偏差并挑战模型以超越它们来评估模型的鲁棒性。该框架由一个偏差放大的训练集和一个困难的反偏差测试集定义，我们使用基于模型的方法从现有数据集中自动提取这些集。

我们的实验表明，识别的反偏差示例对模型来说是自然具有挑战性的，而且，在偏差放大数据上训练的模型在反偏差示例上表现出显著的性能下降，这些下降并未通过常见的提高泛化的方法得到缓解。

在这里，我们将我们的框架应用于Quora问题对数据集（QQP），该数据集由问题对组成，任务是确定问题是否是彼此的释义（具有相同的含义）。

我们的评估框架可以应用于任何现有数据集，甚至是那些被认为是过时的数据集，以测试模型的鲁棒性。我们希望我们的工作将指导开发不依赖于表面偏差和相关性的鲁棒模型。

数据集结构

数据实例

数据实例直接来自QQP（GLUE版本），并重新划分为偏差和反偏差子集。以下是数据集中的一个实例示例： json { "idx": 56, "question1": "How do I buy used car in India?", "question2": "Which used car should I buy in India?", "label": 0 }

数据字段

• idx: 示例在其原始数据划分中的唯一标识符（例如，验证集）
• question1: 在Quora上提出的一个问题
• question2: 在Quora上提出的一个问题
• label: 其中之一 0 和 1（not duplicate 和 duplicate）

数据划分

偏差放大划分需要一种方法来检测数据集中的偏差和反偏差示例。我们发布了使用以下两种方法创建的偏差放大划分：

• 少数示例：我们引入的一种新方法，利用表示学习和聚类来识别反偏差的少数示例（Tu et al., 2020）——那些违背数据集中常见统计模式的示例。
• 部分输入基线：一种常见的方法，用于识别包含数据集中注释伪影的偏差示例，该方法检查仅使用部分输入的模型的性能。如果这些模型成功，它们必然依赖于数据集中的意外或虚假模式。

使用每种方法，我们将原始的训练和测试划分分成偏差和反偏差子集。有关更多详细信息，请参见论文。

少数示例

数据集划分	划分中的实例数量
Train - biased	297735
Train - anti-biased	66111
Validation - biased	32968
Validation - anti-biased	7462

部分输入基线

数据集划分	划分中的实例数量
Train - biased	297735
Train - anti-biased	66111
Validation - biased	33084
Validation - anti-biased	7346

搜集汇总

数据集介绍

构建方式

在自然语言处理领域，模型常依赖数据集中的浅层偏差实现高表现，却难以应对偏离这些统计模式的样本。为促进鲁棒模型的发展，该数据集基于Quora问题对数据集（QQP），采用两种自动检测偏差的方法进行构建。其一为“少数样本法”，借助表示学习与聚类技术，识别出那些违背数据集中常见统计模式的对抗性样本；其二为“部分输入基线法”，通过限制模型仅使用部分输入来发掘标注伪影或虚假关联。由此，原始训练集与验证集被重新划分为偏差增强子集与抗偏差子集，形成具有挑战性的评估框架。

特点

该数据集的核心特点在于其独特的双轨划分结构：针对每种偏差检测方法，均提供偏差增强的训练集与抗偏差的测试集。偏差增强训练集通过放大原有偏差，迫使模型在训练中面对更强的统计信号；而抗偏差测试集则精选出自然难以拟合的样本，用于检验模型是否真正学习了语义特征而非表面相关。实验表明，即便如DeBERTa-large等强大模型，在偏差增强训练后，于抗偏差测试集上的准确率亦会从90%以上骤降至36.8%至63.9%，显著暴露了现有模型对偏差的脆弱依赖。

使用方法

使用该数据集时，需通过HuggingFace的datasets库加载，并选择具体的配置名称，如“minority_examples”或“partial_input”。加载后，可直接通过split参数获取偏差增强训练集（如train.biased）与抗偏差验证集（如validation.anti_biased）进行训练与评估。该框架旨在作为现有基准的补充，用于测试模型在对抗性设置下的鲁棒性，而非替代原始完整数据集。用户可基于此设计实验，对比模型在原始测试集与抗偏差测试集上的表现差异，从而深入诊断其泛化能力。

背景与挑战

背景概述

在自然语言处理领域，模型常依赖数据集中的表面线索（即数据集偏差）取得优异性能，却难以泛化至缺乏此类偏差的样本。为应对这一挑战，Yuval Reif与Roy Schwartz于2023年提出了一种名为“以偏纠偏”的评估框架，并以Quora问题对数据集（QQP）为基底，构建了bias-amplified-splits/qqp数据集。该数据集由希伯来大学团队创建，其核心研究问题在于：通过放大训练数据中的偏差，迫使模型在反偏差测试集上展现真正的鲁棒性。该工作为GLUE基准等经典数据集注入了新的评估维度，推动了鲁棒NLP模型的发展，其影响力体现在对模型表面相关性依赖行为的揭露与度量上。

当前挑战

该数据集面临的核心挑战来自两大层面。其一，在领域问题层面，QQP所解决的是语义等价判别任务，而传统模型易受词汇重叠、句式结构等统计偏差的误导，导致在反偏差样本上性能骤降（如DeBERTa-large在反偏差测试集上准确率从93.0%跌至36.8%）。其二，在构建过程中，如何自动且准确地识别偏差样本与反偏差样本是重大难题。团队需借助表征学习与聚类技术（如少数示例法）或部分输入基线方法，从原始数据中分离出违背常见统计模式的样本，同时确保分拆后的子集既能放大偏差又不失自然语义的完整性，这对算法设计与数据平衡提出了严苛要求。

常用场景

经典使用场景

bias-amplified-splits/qqp 数据集是围绕 Quora 问题对（QQP）任务构建的偏置放大评估框架，其经典使用场景在于评估自然语言理解模型对数据集中表面偏置的依赖程度。通过将原始训练集划分为偏置放大（biased）与反偏置（anti-biased）子集，研究人员可在保持任务语义不变的前提下，检验模型在面对违背统计规律的困难样本时的泛化能力。该框架允许用户从 minority_examples 或 partial_input 两种偏置检测方法中选择，以生成具有挑战性的验证集，从而揭示模型是否真正掌握语义等价判断能力，抑或仅仅利用数据中的虚假关联。

衍生相关工作

该数据集衍生出一系列探索模型鲁棒性提升策略的经典工作。其核心方法 minority_examples 借鉴了表征学习与聚类的思想，启发了后续基于对比学习与反事实数据增强的偏置缓解研究。partial_input 基线方法则推动了部分输入模型在分析标注伪迹方面的应用。此外，该数据集引出的“以偏置制偏置”（Fighting Bias with Bias）理念，直接影响了对抗训练、重加权采样以及鲁棒性正则化等方向的发展，相关论文已在 arXiv 及 ICLR 等平台发表，为构建更可靠的深度语言模型提供了理论基石与实验基准。

数据集最近研究

最新研究方向

在自然语言处理领域，数据集偏差导致的模型泛化能力不足一直是前沿研究的核心挑战。针对Quora问题对（QQP）数据集，最新研究提出了“偏差放大分割”框架，通过自动识别并放大训练集中的统计偏差，同时构建反偏差测试集，系统性地评估模型对表面线索的依赖程度。这一方法颠覆了传统去偏差思路，转而利用偏差本身作为压力测试工具。实验表明，即便如DeBERTa-large等强模型，在反偏差测试集上性能也骤降至36.8%，揭示了现有鲁棒性提升技术的局限性。该工作与近期对模型捷径学习、虚假相关性的广泛关注紧密相连，为构建真正理解语义而非依赖模式匹配的鲁棒NLP系统提供了新的评估范式和基准，对推动可解释、可信赖人工智能的发展具有深远意义。