fahmidiqbal/snli_cf_without_instructions

SNLI-CF Counterfactual Sentence-Only

Dataset Summary

SNLI-CF Counterfactual Sentence-Only is a minimally prompted version of the counterfactually revised Stanford Natural Language Inference (SNLI) examples introduced by Kaushik et al. (ICLR 2020).

The dataset contains counterfactual premise–hypothesis pairs designed to support alternative inference labels.

Each example uses a minimal input format containing the marked premise and hypothesis, followed by a short constrained-output instruction.

Unlike the instruction-based version, this dataset excludes detailed NLI classification rules and label definitions.

The dataset is designed for sentence-only fine-tuning and evaluating generalization to counterfactually revised examples.

Original source: SNLI-CF Kaushik

Dataset Structure

Each example contains:

Column	Description
id	Unique sequential identifier
text	Marked counterfactual premise–hypothesis pair
label	Original counterfactual gold label converted to uppercase

Labels

• ENTAILMENT
• NEUTRAL
• CONTRADICTION

The original NLI label semantics are preserved.

Input Format

Each example follows this template:

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 corresponding counterfactual sentences.

The sentence-only format excludes the detailed inference rules and label descriptions while retaining the output-format instruction.

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 preserved without sampling or deduplication.

Data Preprocessing

The original SNLI-CF data includes the fields idx, premise, hypothesis, label, and type.

The preprocessing pipeline:

1. Reads the original train, validation, and test JSONL files.
2. Selects records where type = cf.
3. Extracts the counterfactual premise and hypothesis.
4. Inserts the sentences inside <p> ... </p> and <h> ... </h> tags.
5. Adds the short label-output instruction.
6. Converts labels to uppercase.
7. Assigns unique sequential IDs.
8. Removes the source idx and type attributes.
9. Saves the examples as JSONL files.

Unique Identifiers

The IDs follow this structure:

cf_train_000001
cf_validation_000001
cf_test_000001

The identifiers are newly assigned during preprocessing.

The IDs and examples match those in the instruction-based counterfactual dataset.

Comparison with Instruction-Based Counterfactual SNLI

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

Both versions contain exactly the same examples, labels, and split ordering.

This enables controlled comparisons between detailed instruction-based fine-tuning and minimal-prompt fine-tuning.

Research Applications

This dataset is intended for:

• Natural language inference classification
• Counterfactual generalization
• Sentence-only fine-tuning
• Robustness evaluation
• Evaluation of sensitivity to semantic changes
• Comparison of different fine-tuning strategies
• Investigation of generalization trade-offs

Limitations

• The dataset contains counterfactual examples only.
• Multiple revisions may originate from the same original example.
• Original source IDs are not retained.
• Direct alignment between original and counterfactual examples requires the source dataset.
• Results should not be reported as performance on the full SNLI benchmark.

Original Dataset and Attribution

The dataset is derived from:

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