fahmidiqbal/snli_cf_with_instructions

SNLI-CF Counterfactual with Instructions

Dataset Summary

SNLI-CF Counterfactual with Instructions is an instruction-formatted version of the counterfactually revised examples from the Stanford Natural Language Inference (SNLI) counterfactual dataset introduced by Kaushik et al. (ICLR 2020).

The dataset contains revised premise–hypothesis pairs that were created to support alternative inference labels.

Each example is formatted using explicit NLI classification instructions, inference-direction rules, label definitions, and a constrained output format.

The dataset is designed for investigating counterfactual robustness, instruction-based fine-tuning, and generalization of language models.

Original source: SNLI-CF Kaushik

Dataset Structure

Each example contains three fields:

Column	Description
id	Unique sequential identifier
text	Full instruction with the counterfactual premise and hypothesis
label	Original counterfactual gold label converted to uppercase

Labels

The dataset uses three labels:

• ENTAILMENT: The hypothesis is definitely true given the premise.
• NEUTRAL: The hypothesis might be true or false given the premise.
• CONTRADICTION: The hypothesis is definitely false given the premise.

Instruction Format

Each example follows this format:

You are a system for natural language inference classification.

Input:
Two marked spans:
- <p>...</p> = PREMISE
- <h>...</h> = HYPOTHESIS

Rules:
- Determine the relationship of the HYPOTHESIS with respect to the PREMISE.
- Use only the information provided in the PREMISE.
- Do NOT assume information that is not stated or logically implied.
- Do NOT reverse the direction of inference.

Labels:
- ENTAILMENT = The hypothesis is definitely true given the premise.
- NEUTRAL = The hypothesis might be true or false given the premise.
- CONTRADICTION = The hypothesis is definitely false given the premise.

Input Pair:
<p> {premise} </p>
<h> {hypothesis} </h>

Output EXACTLY one label: ENTAILMENT, CONTRADICTION, or NEUTRAL.
DO NOT output anything else.

The placeholders {premise} and {hypothesis} are replaced with the counterfactually revised sentences.

The gold label is stored separately in the label field.

Dataset Splits

Split	Examples
Train	6,664
Validation	800
Test	1,600
Total	9,064

Label Distribution

Split	ENTAILMENT	NEUTRAL	CONTRADICTION
Train	2,182	2,241	2,241
Validation	247	255	298
Test	508	554	538

All counterfactual examples are retained.

No additional sampling, deduplication, or shuffling is performed.

Data Preprocessing

The original source contains:

• idx: Original example identifier
• premise: Premise sentence
• hypothesis: Hypothesis sentence
• label: NLI gold label
• type: Original or counterfactual

Preprocessing consists of:

1. Reading train, validation, and test JSONL files.
2. Selecting examples where type = cf.
3. Extracting counterfactual premises and hypotheses.
4. Inserting the sentences inside <p> ... </p> and <h> ... </h> tags.
5. Adding the full NLI classification instructions.
6. Converting the original labels to uppercase.
7. Assigning unique sequential IDs.
8. Removing the original idx and type fields.
9. Saving processed examples as JSONL files.

Unique Identifiers

Each example receives a unique ID:

cf_train_000001
cf_validation_000001
cf_test_000001

The IDs are newly generated and do not preserve the original source identifiers.

Counterfactual Characteristics

The counterfactual examples contain human revisions to the original premise–hypothesis pairs.

The revisions are intended to change the inference relationship while preserving much of the surrounding linguistic context.

The original source contains multiple counterfactual variants for some original examples.

Because the preprocessing removes the source idx attribute, the resulting dataset does not directly encode which counterfactual examples correspond to the same original example.

Relationship to Other Versions

This dataset has a corresponding sentence-only version.

Feature	With Instructions	Sentence-Only
Counterfactual examples	Same	Same
Gold labels	Same	Same
Unique IDs	Same	Same
Full inference instructions	Yes	No
Explicit label definitions	Yes	No
Short output instruction	Yes	Yes

Both versions contain identical examples and labels, allowing controlled comparisons of instruction-based and sentence-only fine-tuning.

Research Applications

The dataset supports:

• Counterfactual generalization
• Natural language inference classification
• Instruction-based fine-tuning
• Robustness evaluation
• Sensitivity to minimal semantic changes
• Generalization across training strategies
• Evaluation of models on revised premise–hypothesis pairs

Limitations

• The dataset contains only counterfactually revised examples, not their original counterparts.
• Some examples share a common original source.
• The original source identifiers are not preserved.
• Direct original-to-counterfactual pair-level analysis requires the original dataset.
• The dataset should not be treated as equivalent to the complete SNLI benchmark.

Original Dataset and Attribution

The counterfactually augmented data was introduced in:

Kaushik, Divyansh, Eduard Hovy, and Zachary C. Lipton. (2020).

Learning the Difference that Makes a Difference with Counterfactually-Augmented Data.

International Conference on Learning Representations (ICLR 2020).

Links:

• SNLI-CF Source Dataset
• Original SNLI Dataset
• Counterfactual Data Paper