Datasets:

Ahren09
/

SARA-QASPER

id stringlengths 1 4	example_id stringlengths 1 3	question stringlengths 12 167	context listlengths 1 168	answer stringlengths 0 1.61k	answer_reformatted stringlengths 25 948 ⌀	choices null	question_type stringclasses 3 values
0	0	What is the seed lexicon?	[ "Experiments ::: Results and Discussion\t Table TABREF23 shows accuracy. As the Random baseline suggests, positive and negative labels were distributed evenly. The Random+Seed baseline made use of the seed lexicon and output the corresponding label (or the reverse of it for negation) if the event's predicate is in ...	a vocabulary of positive and negative predicates that helps determine the polarity score of an event	The seed lexicon is a vocabulary consisting of positive and negative predicates used to determine the polarity score of an event.	null	free_form
1	0	What are the results?	[ "This suggests that the training set of 0.6 million events is sufficiently large for training the models. For comparison, we trained the models with a subset (6,000 events) of the ACP dataset. As the results shown in Table TABREF24 demonstrate, our method is effective when labeled data are small. The result of hype...	Using all data to train: AL -- BiGRU achieved 0.843 accuracy, AL -- BERT achieved 0.863 accuracy, AL+CA+CO -- BiGRU achieved 0.866 accuracy, AL+CA+CO -- BERT achieved 0.835, accuracy, ACP -- BiGRU achieved 0.919 accuracy, ACP -- BERT achived 0.933, accuracy, ACP+AL+CA+CO -- BiGRU achieved 0.917 accuracy, ACP+AL+CA+CO -...	The results show that when using all data to train, BiGRU achieved an accuracy of 0.843 with AL, while BERT achieved 0.863 with AL. For the combined dataset AL+CA+CO, BiGRU achieved 0.866 and BERT achieved 0.835. When using the ACP dataset, BiGRU reached 0.919 and BERT reached 0.933. With the semi-supervised dataset AC...	null	free_form
2	0	How are relations used to propagate polarity?	[ "Related Work\tLearning affective events is closely related to sentiment analysis. Whereas sentiment analysis usually focuses on the polarity of what are described (e.g., movies), we work on how people are typically affected by events. In sentiment analysis, much attention has been paid to compositionality. Word-le...	based on the relation between events, the suggested polarity of one event can determine the possible polarity of the other event	Relations between events are utilized to propagate polarity by allowing the suggested polarity of one event to inform and determine the potential polarity of another event, as demonstrated through discourse relations like Cause and Concession.	null	free_form
3	0	How big is the Japanese data?	[ "Even if $x_2$'s polarity is not known in advance, we can exploit the tendency of $x_1$ and $x_2$ to be of the same polarity (for Cause) or of the reverse polarity (for Concession) although the heuristic is not exempt from counterexamples. We transform this idea into objective functions and train neural network mod...	7000000 pairs of events were extracted from the Japanese Web corpus, 529850 pairs of events were extracted from the ACP corpus	The Japanese web corpus contains 7,000,000 pairs of events, while the ACP corpus consists of 529,850 pairs of events.	null	free_form
4	0	What are labels available in dataset for supervision?	[ "Experiments ::: Results and Discussion\t Table TABREF23 shows accuracy. As the Random baseline suggests, positive and negative labels were distributed evenly. The Random+Seed baseline made use of the seed lexicon and output the corresponding label (or the reverse of it for negation) if the event's predicate is in ...	negative, positive	The labels available in the dataset for supervision are "negative" and "positive," representing the polarity of affective events.	null	extractive
5	0	How big are improvements of supervszed learning results trained on smalled labeled data enhanced with proposed approach copared to basic approach?	[ "Minimally Supervised Learning of Affective Events Using Discourse Relations\tRecognizing affective events that trigger positive or negative sentiment has a wide range of natural language processing applications but remains a challenging problem mainly because the polarity of an event is not necessarily predictable...	3%	Improvements in supervised learning results trained on smaller labeled data enhanced with the proposed approach compared to the basic approach are 3%.	null	free_form
6	0	How does their model learn using mostly raw data?	[ "Minimally Supervised Learning of Affective Events Using Discourse Relations\tRecognizing affective events that trigger positive or negative sentiment has a wide range of natural language processing applications but remains a challenging problem mainly because the polarity of an event is not necessarily predictable...	by exploiting discourse relations to propagate polarity from seed predicates to final sentiment polarity	The model learns using mostly raw data by exploiting discourse relations to efficiently propagate polarity from seed predicates to determine the final sentiment polarity.	null	free_form
7	0	How big is seed lexicon used for training?	[ "Experiments ::: Results and Discussion\t Table TABREF23 shows accuracy. As the Random baseline suggests, positive and negative labels were distributed evenly. The Random+Seed baseline made use of the seed lexicon and output the corresponding label (or the reverse of it for negation) if the event's predicate is in ...	30 words	The seed lexicon used for training consists of 30 words, comprising 15 positive words and 15 negative words.	null	free_form
8	0	How large is raw corpus used for training?	[ "Compared with this method, our discourse relation-based linking of events is much simpler and more intuitive. Some previous studies made use of document structure to understand the sentiment. proposed a sentiment-specific pre-training strategy using unlabeled dialog data (tweet-reply pairs). proposed a method of...	100 million sentences	The raw corpus used for training consists of approximately 100 million sentences.	null	extractive
9	1	Does the paper report macro F1?	["The best model overall is DBMDZ (.520), showing a balanced response on both validation and test se(...TRUNCATED)	Yes	Yes, the paper reports the macro F1 score.	null	yes_no

End of preview. Expand in Data Studio

SARA QASPER (reformatted)

Reformatted QASPER data used by SARA: Selective and Adaptive Retrieval-augmented Generation with Context Compression (ACL 2026, arXiv:2507.05633). Code: Ahren09/SARA.

The SARA Quick Start (python -m src.data.make_qasper_splits) downloads this dataset automatically; you can also load it directly:

from datasets import load_dataset

qa = load_dataset("Ahren09/SARA-QASPER", "qa")                          # train / test
align = load_dataset("Ahren09/SARA-QASPER", "compression_alignment")    # train / validation

Configs

`qa` (default)

One record per answerable QASPER question. train (2,321 rows, 2,321 questions over the official train papers) and test (1,312 rows, official test papers).

Field	Type	Description
`id`	str	Running row index.
`example_id`	str	Paper index — used for leakage-safe document-level train/dev splitting.
`question`	str	The QASPER question.
`context`	list[str]	BM25-ranked paper contexts, each formatted `"Section name\t<text>"`. Retrieval in SARA is built in-memory from this field; no separate index is needed.
`answer`	str	Gold answer derived from the official QASPER annotations.
`answer_reformatted`	str	Fluent rewrite of `answer` (see provenance) — used as the training target and evaluation reference in SARA.
`choices`	null	Unused for QASPER (kept for schema compatibility with multiple-choice datasets).
`question_type`	str	One of `extractive`, `free_form`, `yes_no`, `unanswerable`.

`compression_alignment`

Text snippets from QASPER paper bodies used for the SARA projector-alignment warm-up (the projector learns to reconstruct a document from its semantic compression vector before QA fine-tuning). train (22,111 rows) and validation (200 rows); each record is {"text": "<document text snippet>"}.

Provenance

Derived from allenai/qasper (Dasigi et al., NAACL 2021), released under CC BY 4.0. This derivative is released under the same license with attribution.
context was built from each paper's title/abstract/sections and BM25-ranked per question; BIBREF citation markers were stripped.
answer_reformatted was generated once with gpt-4o-mini (prompted to rewrite the gold answer fluently given the question). It is distributed here — rather than regenerated — because the SARA release adapters were trained against these exact targets and the reported metrics use them as references; regeneration is non-deterministic and would change results.

Checksums (sha256)

2af09d27b9bc6f27e84a3e52db0f4ddd9ba2fe47179f2a29974317e741f8c552  QASPER_train.jsonl
b9ba8096777950b8663c93f58b989bc619c85c14be3c0f091d192c7cf580bdb0  QASPER_test.jsonl
1d6386a84408127a20f01db92f8b9a73ec9499d884d0cab3c2c9f07c4494aa12  QASPER_compression_alignment_train.jsonl
9de81c3cbf42aa4b4001a57762c37f981eb2ae0c5b468c3bfebffe65438f8a0f  QASPER_compression_alignment_dev.jsonl

Citation

@inproceedings{jin2025sara,
  title={SARA: Selective and Adaptive Retrieval-augmented Generation with Context Compression},
  author={Jin, Yiqiao and Sharma, Kartik and Rakesh, Vineeth and Dou, Yingtong and Pan, Menghai and Das, Mahashweta and Kumar, Srijan},
  booktitle={ACL},
  year={2026}
}

@inproceedings{dasigi2021dataset,
  title={A Dataset of Information-Seeking Questions and Answers Anchored in Research Papers},
  author={Dasigi, Pradeep and Lo, Kyle and Beltagy, Iz and Cohan, Arman and Smith, Noah A. and Gardner, Matt},
  booktitle={NAACL},
  year={2021}
}