FlipAware_Triples_smoke

Triples of an anchor sentence, a true paraphrase, and a look alike trap that shares most words but means something different. Built by an automated pipeline and used to train and test the FlipAware embedders.

Leaderboard: FlipAware Leaderboard | Code: GitHub

from datasets import load_dataset

triples = load_dataset("AmareshHebbar/FlipAware_Triples_smoke")
candidates = load_dataset("AmareshHebbar/FlipAware_Triples_smoke", "candidates")
seeds = load_dataset("AmareshHebbar/FlipAware_Triples_smoke", "seeds")
xgen_test = load_dataset("AmareshHebbar/FlipAware_Triples_smoke", "xgen_test")
passage_test = load_dataset("AmareshHebbar/FlipAware_Triples_smoke", "passage_test")

Trap types

• negation: Add or remove a negation so the meaning flips, for example is and is not, can and cannot, always and never.
• numbers: Change one number, amount, date, or unit so the stated fact becomes different.
• role swap: Swap who does what to whom while reusing the same words, for example the dog bit the man and the man bit the dog.
• time order: Reverse the order of two events, or change before into after.
• quantifier: Change a quantifier so the claim covers a different amount, for example some, all, none, most, few.
• hedge: Change how certain the claim is, for example might and will, possibly and definitely, rarely and always.
• entity swap: Replace one specific person, place, organization, or object with a different one of the same kind.
• scope: Move a modifier or negation so it applies to a different part of the sentence, for example only A did B and A did only B.

Contents

Config	Rows	What it holds
default (train)	116	Filtered triples
default (dev)	14	Filtered triples
default (test)	15	Filtered triples
candidates	320	Every generation with its NLI scores, word overlap, and the reason it was kept or rejected
seeds	4000	Source sentences, the public dataset each came from, and the passage around it
xgen_test	38	Test set written by a different model family (microsoft/Phi-4-mini-instruct) on the same held out sentences
passage_test	5	Passage retrieval traps: one sentence inside a real passage is flipped

Splits are made by source sentence, so no anchor in test appears in train or validation.

Pipeline

Step	What happens
1. Seeds	Real English sentences of 8 to 30 words from GooAQ answers and AllNLI premises, deduplicated and tagged when they contain a number
2. Generate	Qwen/Qwen3-4B-Instruct-2507 runs locally with thinking off, sees one worked example per trap type, writes 4 candidates per seed and trap type, and keeps the first that passes the edit check
3. Filter	cross-encoder/nli-deberta-v3-base scores entailment both ways. Kept only when the paraphrase entails the anchor both ways (at least 0.7), the trap does not (below 0.5), and the trap keeps at least 70% of the anchor's words while adding at most 4 new ones
4. Split	80 / 10 / 10 by source sentence
5. Cross generator	microsoft/Phi-4-mini-instruct rewrites the held out test sentences, filtered the same way
6. Passages	Test sentences that came from a multi sentence answer are flipped inside that answer
7. Human	People write and cross check new triples on sentences never shown to any generator

Filter statistics

Trap type	Generated	Skipped	Unparsed	Paraphrase rejected	Trap rejected	Low overlap	Kept	Keep rate
negation	40	0	0	10	1	0	12	30.0%
numbers	40	0	0	6	2	1	22	55.0%
role swap	40	0	0	7	3	0	20	50.0%
time order	40	0	0	4	6	2	25	62.5%
quantifier	40	0	0	5	2	0	7	17.5%
hedge	40	0	0	4	5	0	18	45.0%
entity swap	40	0	0	13	1	0	15	37.5%
scope	40	0	0	3	4	0	26	65.0%
all	320	0	0	52	24	3	145	45.3%

Filter statistics, cross generator

Trap type	Generated	Skipped	Unparsed	Paraphrase rejected	Trap rejected	Low overlap	Kept	Keep rate
negation	15	0	0	7	0	0	3	20.0%
numbers	15	0	0	6	0	0	6	40.0%
role swap	15	0	0	1	2	2	9	60.0%
time order	15	0	0	5	2	1	7	46.7%
quantifier	15	0	0	6	0	0	5	33.3%
hedge	15	0	0	4	7	0	3	20.0%
entity swap	15	0	0	7	0	1	4	26.7%
scope	15	0	0	6	4	2	1	6.7%
all	120	0	0	42	15	6	38	31.7%

Generation prompt

System:

You create test data for embedding models. You always reply with a single JSON object and nothing else.

User:

Trap type: {name}
Rule: {rule}

Example
Source: {ex_anchor}
{"paraphrase": "{ex_para}", "trap": "{ex_trap}"}

Now the real task.
Source: {seed}

paraphrase: the same meaning as the source, with a few words changed. Every fact stays true.
trap: copy the source and change only one to three words, following the rule, so the meaning becomes different.

If the rule cannot apply to this source, reply {"skip": true}
Otherwise reply with the JSON object only, in the same form as the example.

Columns

• default, xgen_test: id, type, anchor, paraphrase, trap
• human_test: adds author, yes, no (check votes)
• candidates: adds generator, nli_para, nli_trap, overlap, kept, reason
• seeds: seed_id, text, source, has_number, context
• passage_test: query, original, twin (the passage with one sentence flipped)

Sources and license

Seed sentences come from GooAQ and AllNLI. Released under CC BY SA 4.0 to match AllNLI.

Limitations

• Generated rows are machine written and machine filtered, so some noise remains.
• English only.

Pipeline code: https://github.com/amareshhebbar/flipaware