fahmidiqbal/ANLI_without_instructions

SNLI R1/R2/R3 Test Sets — Sentence-Only

Dataset Summary

This dataset contains three SNLI-related natural language inference (NLI) test sets, R1, R2, and R3, converted into a minimal, sentence-only prompt format. Each example contains a marked premise–hypothesis pair followed by a constrained output instruction.

Unlike the instruction-based counterpart, this version does not include explicit classification rules, inference-direction guidelines, or definitions of the three labels. The short instruction requesting exactly one label remains part of the prompt.

The three source test sets are kept separate. This repository contains evaluation data only, with no training or validation splits.

The specific definitions or construction procedures for R1, R2, and R3 are not established by the supplied Parquet files. The names should not be interpreted as verified reasoning categories or difficulty levels without additional documentation.

Dataset Files and Splits

Split	Filename	Examples
test_r1	test_r1.jsonl	1,000
test_r2	test_r2.jsonl	1,000
test_r3	test_r3.jsonl	1,200
Total		3,200

All examples from each input file are retained, in their original order. There is no sampling, shuffling, deduplication, or merging of test sets.

Dataset Structure

Each JSONL record contains exactly three fields:

Field	Description
id	Newly generated sequential ID, including the test-set name.
text	Marked premise and hypothesis plus a short output instruction.
label	Gold NLI class name in uppercase.

IDs follow the patterns test_r1_000001, test_r2_000001, and test_r3_000001. IDs match those in the corresponding instruction-based version; they are not original source uid values.

Labels

Source label	Output label
0	ENTAILMENT
1	NEUTRAL
2	CONTRADICTION

The original three-class inference meanings are preserved.

Input Format

Each text field follows this exact template, with {premise} and {hypothesis} replaced by the source sentences:

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

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

The gold answer is stored in the separate label field, not in the input prompt.

Sentence-only means the full NLI instructions are omitted; the short output-format constraint remains.